file stringlengths 6 242 | name stringlengths 1 11 | kernel stringclasses 4 values | input stringlengths 531 14.6k | output stringlengths 176 17.2k |
|---|---|---|---|---|
[407f97a0b39db63cc31f3e635caadfd0,1,784,128,784,128,1,128] | tmp1pe0pk2n | default_function_kernel0 | .visible .entry default_function_kernel0(
.param .u64 default_function_kernel0_param_0,
.param .u64 default_function_kernel0_param_1,
.param .u64 default_function_kernel0_param_2,
.param .u64 default_function_kernel0_param_3
)
{
.reg .pred %p<5>;
.reg .f32 %f<835>;
.reg .b32 %r<127>;
.reg .b64 %rd<32>;
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E20placeholder_d_shared[448];
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E18placeholder_shared[57344];
ld.param.u64 %rd1, [default_function_kernel0_param_0];
ld.param.u64 %rd2, [default_function_kernel0_param_1];
mov.u32 %r13, %tid.x;
mul.lo.s32 %r1, %r13, 3584;
mov.u32 %r123, 0;
mov.f32 %f771, 0f00000000;
cvta.to.global.u64 %rd5, %rd1;
shl.b32 %r17, %r13, 2;
mov.u32 %r18, _ZZ24default_function_kernel0E20placeholder_d_shared;
add.s32 %r19, %r18, %r17;
cvta.to.global.u64 %rd8, %rd2;
for.size.1 i in range(0, 63, 1):
mov.f32 %f(772+i*1), %f771;
$L__BB0_1:
mov.u32 %r124, 0;
bar.sync 0;
mul.lo.s32 %r3, %r123, 112;
add.s32 %r16, %r3, %r13;
mul.wide.s32 %rd6, %r16, 4;
add.s64 %rd7, %rd5, %rd6;
for.size.2 i in range(0, 56, 1):
ld.global.nc.f32 %f(321+i*1), [%rd7+(0+i*8)];
st.shared.f32 [%r19+(0+i*8)], %f(321+i*1);
$L__BB0_2:
mul.lo.s32 %r117, %r123, 112;
shl.b32 %r21, %r124, 1;
add.s32 %r22, %r21, %r13;
mul.hi.s32 %r23, %r22, -1840700269;
add.s32 %r24, %r23, %r22;
shr.u32 %r25, %r24, 31;
shr.s32 %r26, %r24, 6;
add.s32 %r27, %r26, %r25;
mul.lo.s32 %r28, %r27, 112;
sub.s32 %r29, %r22, %r28;
add.s32 %r30, %r29, %r117;
mad.lo.s32 %r31, %r27, 784, %r30;
mul.wide.s32 %rd9, %r31, 4;
add.s64 %rd10, %rd8, %rd9;
ld.global.nc.f32 %f377, [%rd10];
shl.b32 %r32, %r22, 2;
mov.u32 %r33, _ZZ24default_function_kernel0E18placeholder_shared;
add.s32 %r34, %r33, %r32;
for.size.14 i in range(0, 7, 1):
st.shared.f32 [%r34+(0+i*8)], %f(377+i*1);
add.s32 %r(35+i*10), %r22, (2+i*2);
mul.hi.s32 %r(36+i*10), %r(35+i*10), -1840700269;
add.s32 %r(37+i*10), %r(36+i*10), %r(35+i*10);
shr.u32 %r(38+i*10), %r(37+i*10), 31;
shr.s32 %r(39+i*10), %r(37+i*10), 6;
add.s32 %r(40+i*10), %r(39+i*10), %r(38+i*10);
mul.lo.s32 %r(41+i*10), %r(40+i*10), 112;
sub.s32 %r(42+i*10), %r(35+i*10), %r(41+i*10);
add.s32 %r(43+i*10), %r(42+i*10), %r117;
mad.lo.s32 %r(44+i*10), %r(40+i*10), 784, %r(43+i*10);
mul.wide.s32 %rd(11+i*2), %r(44+i*10), 4;
add.s64 %rd(12+i*2), %rd8, %rd(11+i*2);
ld.global.nc.f32 %f(378+i*1), [%rd(12+i*2)+0];
st.shared.f32 [%r34+56], %f384;
add.s32 %r124, %r124, 8;
setp.ne.s32 %p1, %r124, 7168;
@%p1 bra $L__BB0_2;
bar.sync 0;
mov.u32 %r125, 0;
$L__BB0_4:
mov.u32 %r126, 0;
$L__BB0_5:
mul.lo.s32 %r121, %r125, 56;
mov.u32 %r119, _ZZ24default_function_kernel0E18placeholder_shared;
mov.u32 %r118, _ZZ24default_function_kernel0E20placeholder_d_shared;
for.size.3 i in range(0, 2, 1):
add.s32 %r(107+i*4), %r(126+i*-19), %r(121+i*-120);
shl.b32 %r(108+i*4), %r(107+i*4), 2;
add.s32 %r(110+i*4), %r(118+i*1), %r(108+i*4);
for.size.1 i in range(0, 2, 1):
ld.shared.f32 %f(385+i*1), [%r(114+i*-4)+0];
for.size.2 i in range(0, 2, 1):
fma.rn.f32 %f(834+i*-1), %f386, %f(385+i*2), %f(834+i*-1);
ld.shared.f32 %f(387+i*1), [%r114+(28672+i*-28224)];
for.size.3 i in range(0, 30, 1):
for.size.2 j in range(0, 2, 1):
fma.rn.f32 %f(832+i*-2+j*-1), %f386, %f(388+i*2+j*1), %f(832+i*-2+j*-1);
ld.shared.f32 %f(389+i*2+j*1), [%r114+(29120+i*448+j*-28224)];
fma.rn.f32 %f772, %f386, %f448, %f772;
ld.shared.f32 %f449, [%r114+42560];
fma.rn.f32 %f771, %f386, %f449, %f771;
add.s32 %r126, %r126, 1;
setp.ne.s32 %p2, %r126, 56;
@%p2 bra $L__BB0_5;
add.s32 %r125, %r125, 1;
setp.lt.u32 %p3, %r125, 2;
@%p3 bra $L__BB0_4;
add.s32 %r123, %r123, 1;
setp.lt.u32 %p4, %r123, 7;
@%p4 bra $L__BB0_1;
ld.param.u64 %rd31, [default_function_kernel0_param_2];
ld.param.u64 %rd30, [default_function_kernel0_param_3];
shl.b32 %r116, %r13, 5;
cvta.to.global.u64 %rd25, %rd30;
mul.wide.s32 %rd26, %r116, 4;
add.s64 %rd27, %rd25, %rd26;
ld.global.nc.f32 %f450, [%rd27];
add.f32 %f451, %f834, %f450;
mov.f32 %f452, 0f00000000;
max.f32 %f453, %f451, %f452;
cvta.to.global.u64 %rd28, %rd31;
add.s64 %rd29, %rd28, %rd26;
for.size.5 i in range(0, 31, 1):
for.size.4 j in range(0, 2, 1):
st.global.f32 [%rd29+(0+i*4+j*256)], %f(453+i*6+j*3);
ld.global.nc.f32 %f(454+i*6+j*3), [%rd27+(256+i*4+j*-252)];
add.f32 %f(455+i*6+j*3), %f(833+i*-2+j*-1), %f(454+i*6+j*3);
max.f32 %f(456+i*6+j*3), %f(455+i*6+j*3), %f452;
st.global.f32 [%rd29+124], %f639;
ld.global.nc.f32 %f640, [%rd27+380];
add.f32 %f641, %f771, %f640;
max.f32 %f642, %f641, %f452;
st.global.f32 [%rd29+380], %f642;
ret;
}
| extern "C" __global__ void default_function_kernel0(float* __restrict__ placeholder, float* __restrict__ placeholder1, float* __restrict__ T_relu, float* __restrict__ placeholder2) {
float T_dense[64];
__shared__ float placeholder_d_shared[112];
__shared__ float placeholder_shared[14336];
T_dense[0] = 0.000000e+00f;
T_dense[32] = 0.000000e+00f;
T_dense[1] = 0.000000e+00f;
T_dense[33] = 0.000000e+00f;
T_dense[2] = 0.000000e+00f;
T_dense[34] = 0.000000e+00f;
T_dense[3] = 0.000000e+00f;
T_dense[35] = 0.000000e+00f;
T_dense[4] = 0.000000e+00f;
T_dense[36] = 0.000000e+00f;
T_dense[5] = 0.000000e+00f;
T_dense[37] = 0.000000e+00f;
T_dense[6] = 0.000000e+00f;
T_dense[38] = 0.000000e+00f;
T_dense[7] = 0.000000e+00f;
T_dense[39] = 0.000000e+00f;
T_dense[8] = 0.000000e+00f;
T_dense[40] = 0.000000e+00f;
T_dense[9] = 0.000000e+00f;
T_dense[41] = 0.000000e+00f;
T_dense[10] = 0.000000e+00f;
T_dense[42] = 0.000000e+00f;
T_dense[11] = 0.000000e+00f;
T_dense[43] = 0.000000e+00f;
T_dense[12] = 0.000000e+00f;
T_dense[44] = 0.000000e+00f;
T_dense[13] = 0.000000e+00f;
T_dense[45] = 0.000000e+00f;
T_dense[14] = 0.000000e+00f;
T_dense[46] = 0.000000e+00f;
T_dense[15] = 0.000000e+00f;
T_dense[47] = 0.000000e+00f;
T_dense[16] = 0.000000e+00f;
T_dense[48] = 0.000000e+00f;
T_dense[17] = 0.000000e+00f;
T_dense[49] = 0.000000e+00f;
T_dense[18] = 0.000000e+00f;
T_dense[50] = 0.000000e+00f;
T_dense[19] = 0.000000e+00f;
T_dense[51] = 0.000000e+00f;
T_dense[20] = 0.000000e+00f;
T_dense[52] = 0.000000e+00f;
T_dense[21] = 0.000000e+00f;
T_dense[53] = 0.000000e+00f;
T_dense[22] = 0.000000e+00f;
T_dense[54] = 0.000000e+00f;
T_dense[23] = 0.000000e+00f;
T_dense[55] = 0.000000e+00f;
T_dense[24] = 0.000000e+00f;
T_dense[56] = 0.000000e+00f;
T_dense[25] = 0.000000e+00f;
T_dense[57] = 0.000000e+00f;
T_dense[26] = 0.000000e+00f;
T_dense[58] = 0.000000e+00f;
T_dense[27] = 0.000000e+00f;
T_dense[59] = 0.000000e+00f;
T_dense[28] = 0.000000e+00f;
T_dense[60] = 0.000000e+00f;
T_dense[29] = 0.000000e+00f;
T_dense[61] = 0.000000e+00f;
T_dense[30] = 0.000000e+00f;
T_dense[62] = 0.000000e+00f;
T_dense[31] = 0.000000e+00f;
T_dense[63] = 0.000000e+00f;
for (int k_outer_outer = 0; k_outer_outer < 7; ++k_outer_outer) {
__syncthreads();
int index0 = (int)threadIdx.x;
int index1 = k_outer_outer * 112 + (int)threadIdx.x;
placeholder_d_shared[index0] = placeholder[index1];
placeholder_d_shared[index0 + 2] = placeholder[index1 + 2];
placeholder_d_shared[index0 + 4] = placeholder[index1 + 4];
placeholder_d_shared[index0 + 6] = placeholder[index1 + 6];
placeholder_d_shared[index0 + 8] = placeholder[index1 + 8];
placeholder_d_shared[index0 + 10] = placeholder[index1 + 10];
placeholder_d_shared[index0 + 12] = placeholder[index1 + 12];
placeholder_d_shared[index0 + 14] = placeholder[index1 + 14];
placeholder_d_shared[index0 + 16] = placeholder[index1 + 16];
placeholder_d_shared[index0 + 18] = placeholder[index1 + 18];
placeholder_d_shared[index0 + 20] = placeholder[index1 + 20];
placeholder_d_shared[index0 + 22] = placeholder[index1 + 22];
placeholder_d_shared[index0 + 24] = placeholder[index1 + 24];
placeholder_d_shared[index0 + 26] = placeholder[index1 + 26];
placeholder_d_shared[index0 + 28] = placeholder[index1 + 28];
placeholder_d_shared[index0 + 30] = placeholder[index1 + 30];
placeholder_d_shared[index0 + 32] = placeholder[index1 + 32];
placeholder_d_shared[index0 + 34] = placeholder[index1 + 34];
placeholder_d_shared[index0 + 36] = placeholder[index1 + 36];
placeholder_d_shared[index0 + 38] = placeholder[index1 + 38];
placeholder_d_shared[index0 + 40] = placeholder[index1 + 40];
placeholder_d_shared[index0 + 42] = placeholder[index1 + 42];
placeholder_d_shared[index0 + 44] = placeholder[index1 + 44];
placeholder_d_shared[index0 + 46] = placeholder[index1 + 46];
placeholder_d_shared[index0 + 48] = placeholder[index1 + 48];
placeholder_d_shared[index0 + 50] = placeholder[index1 + 50];
placeholder_d_shared[index0 + 52] = placeholder[index1 + 52];
placeholder_d_shared[index0 + 54] = placeholder[index1 + 54];
placeholder_d_shared[index0 + 56] = placeholder[index1 + 56];
placeholder_d_shared[index0 + 58] = placeholder[index1 + 58];
placeholder_d_shared[index0 + 60] = placeholder[index1 + 60];
placeholder_d_shared[index0 + 62] = placeholder[index1 + 62];
placeholder_d_shared[index0 + 64] = placeholder[index1 + 64];
placeholder_d_shared[index0 + 66] = placeholder[index1 + 66];
placeholder_d_shared[index0 + 68] = placeholder[index1 + 68];
placeholder_d_shared[index0 + 70] = placeholder[index1 + 70];
placeholder_d_shared[index0 + 72] = placeholder[index1 + 72];
placeholder_d_shared[index0 + 74] = placeholder[index1 + 74];
placeholder_d_shared[index0 + 76] = placeholder[index1 + 76];
placeholder_d_shared[index0 + 78] = placeholder[index1 + 78];
placeholder_d_shared[index0 + 80] = placeholder[index1 + 80];
placeholder_d_shared[index0 + 82] = placeholder[index1 + 82];
placeholder_d_shared[index0 + 84] = placeholder[index1 + 84];
placeholder_d_shared[index0 + 86] = placeholder[index1 + 86];
placeholder_d_shared[index0 + 88] = placeholder[index1 + 88];
placeholder_d_shared[index0 + 90] = placeholder[index1 + 90];
placeholder_d_shared[index0 + 92] = placeholder[index1 + 92];
placeholder_d_shared[index0 + 94] = placeholder[index1 + 94];
placeholder_d_shared[index0 + 96] = placeholder[index1 + 96];
placeholder_d_shared[index0 + 98] = placeholder[index1 + 98];
placeholder_d_shared[index0 + 100] = placeholder[index1 + 100];
placeholder_d_shared[index0 + 102] = placeholder[index1 + 102];
placeholder_d_shared[index0 + 104] = placeholder[index1 + 104];
placeholder_d_shared[index0 + 106] = placeholder[index1 + 106];
placeholder_d_shared[index0 + 108] = placeholder[index1 + 108];
placeholder_d_shared[index0 + 110] = placeholder[index1 + 110];
for (int ax0_ax1_fused_outer_outer = 0; ax0_ax1_fused_outer_outer < 7168; ++ax0_ax1_fused_outer_outer) {
int index2 = ax0_ax1_fused_outer_outer * 2 + (int)threadIdx.x;
int index3 = (ax0_ax1_fused_outer_outer * 2 + (int)threadIdx.x) / 112 * 784 + k_outer_outer * 112 + (ax0_ax1_fused_outer_outer * 2 + (int)threadIdx.x) % 112;
placeholder_shared[index2] = placeholder1[index3];
}
__syncthreads();
for (int k_outer_inner = 0; k_outer_inner < 2; ++k_outer_inner) {
for (int k_inner = 0; k_inner < 56; ++k_inner) {
int index4 = k_outer_inner * 56 + k_inner;
int index5 = (int)threadIdx.x * 3584 + k_outer_inner * 56 + k_inner;
T_dense[0] = T_dense[0] + placeholder_d_shared[index4] * placeholder_shared[index5];
T_dense[32] = T_dense[32] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 7168]);
T_dense[1] = T_dense[1] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 112]);
T_dense[33] = T_dense[33] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 7280]);
T_dense[2] = T_dense[2] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 224]);
T_dense[34] = T_dense[34] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 7392]);
T_dense[3] = T_dense[3] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 336]);
T_dense[35] = T_dense[35] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 7504]);
T_dense[4] = T_dense[4] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 448]);
T_dense[36] = T_dense[36] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 7616]);
T_dense[5] = T_dense[5] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 560]);
T_dense[37] = T_dense[37] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 7728]);
T_dense[6] = T_dense[6] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 672]);
T_dense[38] = T_dense[38] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 7840]);
T_dense[7] = T_dense[7] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 784]);
T_dense[39] = T_dense[39] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 7952]);
T_dense[8] = T_dense[8] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 896]);
T_dense[40] = T_dense[40] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 8064]);
T_dense[9] = T_dense[9] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1008]);
T_dense[41] = T_dense[41] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 8176]);
T_dense[10] = T_dense[10] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1120]);
T_dense[42] = T_dense[42] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 8288]);
T_dense[11] = T_dense[11] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1232]);
T_dense[43] = T_dense[43] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 8400]);
T_dense[12] = T_dense[12] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1344]);
T_dense[44] = T_dense[44] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 8512]);
T_dense[13] = T_dense[13] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1456]);
T_dense[45] = T_dense[45] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 8624]);
T_dense[14] = T_dense[14] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1568]);
T_dense[46] = T_dense[46] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 8736]);
T_dense[15] = T_dense[15] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1680]);
T_dense[47] = T_dense[47] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 8848]);
T_dense[16] = T_dense[16] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1792]);
T_dense[48] = T_dense[48] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 8960]);
T_dense[17] = T_dense[17] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1904]);
T_dense[49] = T_dense[49] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 9072]);
T_dense[18] = T_dense[18] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 2016]);
T_dense[50] = T_dense[50] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 9184]);
T_dense[19] = T_dense[19] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 2128]);
T_dense[51] = T_dense[51] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 9296]);
T_dense[20] = T_dense[20] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 2240]);
T_dense[52] = T_dense[52] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 9408]);
T_dense[21] = T_dense[21] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 2352]);
T_dense[53] = T_dense[53] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 9520]);
T_dense[22] = T_dense[22] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 2464]);
T_dense[54] = T_dense[54] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 9632]);
T_dense[23] = T_dense[23] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 2576]);
T_dense[55] = T_dense[55] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 9744]);
T_dense[24] = T_dense[24] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 2688]);
T_dense[56] = T_dense[56] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 9856]);
T_dense[25] = T_dense[25] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 2800]);
T_dense[57] = T_dense[57] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 9968]);
T_dense[26] = T_dense[26] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 2912]);
T_dense[58] = T_dense[58] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 10080]);
T_dense[27] = T_dense[27] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 3024]);
T_dense[59] = T_dense[59] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 10192]);
T_dense[28] = T_dense[28] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 3136]);
T_dense[60] = T_dense[60] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 10304]);
T_dense[29] = T_dense[29] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 3248]);
T_dense[61] = T_dense[61] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 10416]);
T_dense[30] = T_dense[30] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 3360]);
T_dense[62] = T_dense[62] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 10528]);
T_dense[31] = T_dense[31] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 3472]);
T_dense[63] = T_dense[63] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 10640]);
}
}
}
for (int ax1_inner = 0; ax1_inner < 32; ++ax1_inner) {
int index6 = (int)threadIdx.x * 32 + ax1_inner;
int index7 = ax1_inner;
T_relu[index6] = max(T_dense[index7] + placeholder2[index6], 0.000000e+00f);
T_relu[index6 + 64] = max(T_dense[index7 + 32] + placeholder2[index6 + 64], 0.000000e+00f);
}
} |
[407f97a0b39db63cc31f3e635caadfd0,1,784,128,784,128,1,128] | tmp_pu9j6an | default_function_kernel0 | .visible .entry default_function_kernel0(
.param .u64 default_function_kernel0_param_0,
.param .u64 default_function_kernel0_param_1,
.param .u64 default_function_kernel0_param_2,
.param .u64 default_function_kernel0_param_3
)
{
.reg .pred %p<6>;
.reg .f32 %f<82>;
.reg .b32 %r<20>;
.reg .b64 %rd<27>;
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E20placeholder_d_shared[4];
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E18placeholder_shared[64];
ld.param.u64 %rd12, [default_function_kernel0_param_0];
ld.param.u64 %rd13, [default_function_kernel0_param_1];
ld.param.u64 %rd10, [default_function_kernel0_param_2];
ld.param.u64 %rd11, [default_function_kernel0_param_3];
cvta.to.global.u64 %rd14, %rd13;
mov.u32 %r1, %tid.x;
shl.b32 %r3, %r1, 2;
mov.u32 %r11, _ZZ24default_function_kernel0E20placeholder_d_shared;
add.s32 %r2, %r11, %r3;
shl.b32 %r12, %r1, 4;
mov.u32 %r13, _ZZ24default_function_kernel0E18placeholder_shared;
add.s32 %r4, %r13, %r12;
mov.u32 %r5, %ctaid.x;
mul.lo.s32 %r14, %r1, 3136;
mad.lo.s32 %r15, %r5, 12544, %r14;
mul.wide.s32 %rd15, %r15, 4;
add.s64 %rd16, %rd14, %rd15;
add.s64 %rd26, %rd16, 6272;
mul.lo.s32 %r16, %r1, 784;
cvta.to.global.u64 %rd17, %rd12;
mul.wide.s32 %rd18, %r16, 4;
add.s64 %rd19, %rd17, %rd18;
add.s64 %rd25, %rd19, 8;
mov.u32 %r19, 0;
mov.f32 %f78, 0f00000000;
for.size.1 i in range(0, 3, 1):
mov.f32 %f(79+i*1), %f78;
$L__BB0_1:
bar.sync 0;
add.s64 %rd5, %rd25, -8;
setp.gt.s32 %p1, %r1, 0;
@%p1 bra $L__BB0_3;
ld.global.nc.f32 %f25, [%rd5];
st.shared.f32 [%r2], %f25;
$L__BB0_3:
add.s64 %rd6, %rd26, -6272;
for.size.1 i in range(0, 4, 1):
ld.global.nc.f32 %f(26+i*1), [%rd26+(3136+i*-3136)];
st.shared.v4.f32 [%r4], {%f29, %f28, %f27, %f26};
bar.sync 0;
ld.shared.f32 %f30, [%r4];
ld.shared.f32 %f31, [_ZZ24default_function_kernel0E20placeholder_d_shared];
for.size.2 i in range(0, 2, 1):
fma.rn.f32 %f(5+i*1), %f31, %f(30+i*2), %f(81+i*-1);
ld.shared.f32 %f(32+i*1), [%r4+(4+i*4)];
fma.rn.f32 %f7, %f31, %f33, %f79;
ld.shared.f32 %f34, [%r4+12];
fma.rn.f32 %f8, %f31, %f34, %f78;
bar.sync 0;
@%p1 bra $L__BB0_5;
ld.global.nc.f32 %f35, [%rd5+4];
st.shared.f32 [%r2], %f35;
$L__BB0_5:
for.size.1 i in range(0, 4, 1):
ld.global.nc.f32 %f(36+i*1), [%rd6+(9412+i*-3136)];
st.shared.v4.f32 [%r4], {%f39, %f38, %f37, %f36};
bar.sync 0;
ld.shared.f32 %f40, [%r4];
ld.shared.f32 %f41, [_ZZ24default_function_kernel0E20placeholder_d_shared];
for.size.2 i in range(0, 2, 1):
fma.rn.f32 %f(9+i*1), %f41, %f(40+i*2), %f(5+i*1);
ld.shared.f32 %f(42+i*1), [%r4+(4+i*4)];
fma.rn.f32 %f11, %f41, %f43, %f7;
ld.shared.f32 %f44, [%r4+12];
fma.rn.f32 %f12, %f41, %f44, %f8;
bar.sync 0;
@%p1 bra $L__BB0_7;
ld.global.nc.f32 %f45, [%rd5+8];
st.shared.f32 [%r2], %f45;
$L__BB0_7:
for.size.1 i in range(0, 4, 1):
ld.global.nc.f32 %f(46+i*1), [%rd6+(9416+i*-3136)];
st.shared.v4.f32 [%r4], {%f49, %f48, %f47, %f46};
bar.sync 0;
ld.shared.f32 %f50, [%r4];
ld.shared.f32 %f51, [_ZZ24default_function_kernel0E20placeholder_d_shared];
for.size.2 i in range(0, 2, 1):
fma.rn.f32 %f(13+i*1), %f51, %f(50+i*2), %f(9+i*1);
ld.shared.f32 %f(52+i*1), [%r4+(4+i*4)];
fma.rn.f32 %f15, %f51, %f53, %f11;
ld.shared.f32 %f54, [%r4+12];
fma.rn.f32 %f16, %f51, %f54, %f12;
bar.sync 0;
@%p1 bra $L__BB0_9;
ld.global.nc.f32 %f55, [%rd5+12];
st.shared.f32 [%r2], %f55;
$L__BB0_9:
for.size.1 i in range(0, 4, 1):
ld.global.nc.f32 %f(56+i*1), [%rd6+(9420+i*-3136)];
st.shared.v4.f32 [%r4], {%f59, %f58, %f57, %f56};
bar.sync 0;
ld.shared.f32 %f60, [%r4];
ld.shared.f32 %f61, [_ZZ24default_function_kernel0E20placeholder_d_shared];
for.size.2 i in range(0, 2, 1):
fma.rn.f32 %f(81+i*-1), %f61, %f(60+i*2), %f(13+i*1);
ld.shared.f32 %f(62+i*1), [%r4+(4+i*4)];
fma.rn.f32 %f79, %f61, %f63, %f15;
ld.shared.f32 %f64, [%r4+12];
fma.rn.f32 %f78, %f61, %f64, %f16;
for.size.1 i in range(0, 2, 1):
add.s64 %rd(26+i*-1), %rd(26+i*-1), 16;
add.s32 %r19, %r19, 4;
setp.ne.s32 %p5, %r19, 784;
@%p5 bra $L__BB0_1;
cvta.to.global.u64 %rd20, %rd11;
shl.b32 %r17, %r5, 4;
add.s32 %r18, %r17, %r3;
mul.wide.s32 %rd21, %r18, 4;
add.s64 %rd22, %rd20, %rd21;
ld.global.nc.f32 %f65, [%rd22];
add.f32 %f66, %f81, %f65;
mov.f32 %f67, 0f00000000;
max.f32 %f68, %f66, %f67;
cvta.to.global.u64 %rd23, %rd10;
add.s64 %rd24, %rd23, %rd21;
for.size.4 i in range(0, 3, 1):
st.global.f32 [%rd24+(0+i*4)], %f(68+i*3);
ld.global.nc.f32 %f(69+i*3), [%rd22+(4+i*4)];
add.f32 %f(70+i*3), %f(80+i*-1), %f(69+i*3);
max.f32 %f(71+i*3), %f(70+i*3), %f67;
st.global.f32 [%rd24+12], %f77;
ret;
}
| extern "C" __global__ void default_function_kernel0(float* __restrict__ placeholder, float* __restrict__ placeholder1, float* __restrict__ T_relu, float* __restrict__ placeholder2) {
float T_dense[4];
__shared__ float placeholder_d_shared[1];
__shared__ float placeholder_shared[16];
T_dense[0] = 0.000000e+00f;
T_dense[1] = 0.000000e+00f;
T_dense[2] = 0.000000e+00f;
T_dense[3] = 0.000000e+00f;
for (int k_outer_outer = 0; k_outer_outer < 784; ++k_outer_outer) {
__syncthreads();
if ((int)threadIdx.x < 1) {
int index0 = (int)threadIdx.x;
int index1 = (int)threadIdx.x * 784 + k_outer_outer;
placeholder_d_shared[index0] = placeholder[index1];
}
int4 _1 = make_int4((int)blockIdx.x * 12544 + (int)threadIdx.x * 3136 + k_outer_outer+784*0, (int)blockIdx.x * 12544 + (int)threadIdx.x * 3136 + k_outer_outer+784*1, (int)blockIdx.x * 12544 + (int)threadIdx.x * 3136 + k_outer_outer+784*2, (int)blockIdx.x * 12544 + (int)threadIdx.x * 3136 + k_outer_outer+784*3);
int index2 = _1.x;
int index3 = _1.y;
int index4 = _1.z;
int index5 = _1.w;
((float4*)(placeholder_shared + (int)threadIdx.x * 4))[0] = make_float4(placeholder1[index2],placeholder1[index3],placeholder1[index4],placeholder1[index5]);
__syncthreads();
int index0 = (int)threadIdx.x;
T_dense[0] = T_dense[0] + placeholder_d_shared[0] * placeholder_shared[index0 * 4];
int index6 = (int)threadIdx.x * 4 + 1;
T_dense[1] = T_dense[1] + placeholder_d_shared[0] * placeholder_shared[index6];
int index7 = (int)threadIdx.x * 4 + 2;
T_dense[2] = T_dense[2] + placeholder_d_shared[0] * placeholder_shared[index7];
int index8 = (int)threadIdx.x * 4 + 3;
T_dense[3] = T_dense[3] + placeholder_d_shared[0] * placeholder_shared[index8];
}
for (int ax1_inner = 0; ax1_inner < 4; ++ax1_inner) {
int index9 = (int)blockIdx.x * 16 + (int)threadIdx.x * 4 + ax1_inner;
int index10 = ax1_inner;
T_relu[index9] = max(T_dense[index10] + placeholder2[index9], 0.000000e+00f);
}
} |
[407f97a0b39db63cc31f3e635caadfd0,1,784,128,784,128,1,128] | tmp56uqx0tl | default_function_kernel0 | .visible .entry default_function_kernel0(
.param .u64 default_function_kernel0_param_0,
.param .u64 default_function_kernel0_param_1,
.param .u64 default_function_kernel0_param_2,
.param .u64 default_function_kernel0_param_3
)
{
.reg .pred %p<3>;
.reg .f32 %f<67>;
.reg .b32 %r<91>;
.reg .b64 %rd<51>;
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E20placeholder_d_shared[28];
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E18placeholder_shared[112];
ld.param.u64 %rd20, [default_function_kernel0_param_0];
ld.param.u64 %rd21, [default_function_kernel0_param_1];
ld.param.u64 %rd22, [default_function_kernel0_param_2];
ld.param.u64 %rd19, [default_function_kernel0_param_3];
cvta.to.global.u64 %rd50, %rd21;
cvta.to.global.u64 %rd2, %rd22;
mov.u32 %r1, %tid.x;
for.size.1 i in range(0, 2, 1):
shl.b32 %r(9+i*1), %r1, (2+i*2);
mov.u32 %r11, _ZZ24default_function_kernel0E20placeholder_d_shared;
add.s32 %r2, %r11, %r10;
mov.u32 %r12, _ZZ24default_function_kernel0E18placeholder_shared;
add.s32 %r3, %r12, %r9;
add.s32 %r13, %r1, 6;
mul.hi.s32 %r14, %r13, -1840700269;
add.s32 %r15, %r14, %r13;
shr.u32 %r16, %r15, 31;
shr.s32 %r17, %r15, 2;
add.s32 %r18, %r17, %r16;
mul.lo.s32 %r19, %r18, 7;
sub.s32 %r20, %r13, %r19;
for.size.6 i in range(0, 3, 1):
add.s32 %r(21+i*6), %r1, (8+i*2);
mul.hi.s32 %r(22+i*6), %r(21+i*6), -1840700269;
add.s32 %r(23+i*6), %r(22+i*6), %r(21+i*6);
shr.u32 %r(24+i*6), %r(23+i*6), 31;
shr.s32 %r(25+i*6), %r(23+i*6), 2;
add.s32 %r(26+i*6), %r(25+i*6), %r(24+i*6);
for.size.6 i in range(0, 6, 1):
add.s32 %r(39+i*6), %r1, (16+i*2);
mul.hi.s32 %r(40+i*6), %r(39+i*6), -1840700269;
add.s32 %r(41+i*6), %r(40+i*6), %r(39+i*6);
shr.u32 %r(42+i*6), %r(41+i*6), 31;
shr.s32 %r(43+i*6), %r(41+i*6), 2;
add.s32 %r(44+i*6), %r(43+i*6), %r(42+i*6);
mad.lo.s32 %r4, %r1, 28, %r12;
mov.u32 %r5, %ctaid.x;
mul.lo.s32 %r75, %r5, 3136;
add.s32 %r76, %r1, %r75;
for.size.3 i in range(0, 3, 1):
mad.lo.s32 %r(77+i*1), %r(74+i*-6), 784, %r76;
mul.wide.s32 %rd(23+i*1), %r(77+i*1), 4;
add.s64 %rd(3+i*1), %rd(23+i*1), (20+i*-8);
add.s32 %r80, %r20, %r75;
for.size.2 i in range(0, 2, 1):
mad.lo.s32 %r(81+i*1), %r(56+i*-6), 784, %r(80+i*-4);
mul.wide.s32 %rd(6+i*20), %r(81+i*1), 4;
for.size.4 i in range(0, 2, 1):
for.size.3 j in range(0, 2, 1):
add.s64 %rd(7+i*2+j*1), %rd(26+i*2+j*1), (16+i*4+j*-8);
mad.lo.s32 %r(83+i*2+j*1), %r(44+i*-12+j*-6), 784, %r76;
mul.wide.s32 %rd(27+i*2+j*1), %r(83+i*2+j*1), 4;
add.s64 %rd11, %rd30, 4;
mad.lo.s32 %r87, %r18, 784, %r80;
mul.wide.s32 %rd12, %r87, 4;
cvta.to.global.u64 %rd31, %rd20;
mul.wide.s32 %rd32, %r9, 4;
add.s64 %rd33, %rd31, %rd32;
add.s64 %rd49, %rd33, 8;
mul.wide.s32 %rd14, %r76, 4;
mov.f32 %f65, 0f00000000;
mov.u32 %r90, 0;
mov.f32 %f66, %f65;
$L__BB0_1:
bar.sync 0;
for.size.2 i in range(0, 3, 1):
ld.global.nc.f32 %f(7+i*1), [%rd49+(-8+i*4)];
st.shared.f32 [%r2+(0+i*4)], %f(7+i*1);
setp.gt.s32 %p1, %r1, 0;
@%p1 bra $L__BB0_3;
ld.global.nc.f32 %f10, [%rd49+4];
st.shared.f32 [%r2+12], %f10;
$L__BB0_3:
add.s64 %rd34, %rd50, %rd14;
for.size.2 i in range(0, 3, 1):
ld.global.nc.f32 %f(11+i*1), [%rd34+(0+i*8)];
st.shared.f32 [%r3+(0+i*8)], %f(11+i*1);
for.size.3 i in range(0, 4, 1):
add.s64 %rd(35+i*1), %rd50, %rd(12+i*-1);
ld.global.nc.f32 %f(14+i*1), [%rd(35+i*1)+0];
st.shared.f32 [%r3+(24+i*8)], %f(14+i*1);
ld.global.nc.f32 %f18, [%rd34+6272];
for.size.3 i in range(0, 6, 1):
st.shared.f32 [%r3+(56+i*8)], %f(18+i*1);
add.s64 %rd(39+i*1), %rd50, %rd(8+i*-1);
ld.global.nc.f32 %f(19+i*1), [%rd(39+i*1)+0];
st.shared.f32 [%r3+104], %f24;
bar.sync 0;
for.size.5 i in range(0, 2, 1):
ld.shared.f32 %f(25+i*5), [%r4+(0+i*4)];
ld.shared.f32 %f(26+i*5), [_ZZ24default_function_kernel0E20placeholder_d_shared+(0+i*4)];
fma.rn.f32 %f(27+i*5), %f(26+i*5), %f(25+i*5), %f(65+i*-38);
ld.shared.f32 %f(28+i*5), [%r4+(56+i*4)];
fma.rn.f32 %f(29+i*5), %f(26+i*5), %f(28+i*5), %f(66+i*-37);
for.size.5 i in range(0, 4, 1):
ld.shared.f32 %f(35+i*5), [%r4+(8+i*4)];
ld.shared.f32 %f(36+i*5), [_ZZ24default_function_kernel0E20placeholder_d_shared+(8+i*4)];
fma.rn.f32 %f(37+i*5), %f(36+i*5), %f(35+i*5), %f(32+i*5);
ld.shared.f32 %f(38+i*5), [%r4+(64+i*4)];
fma.rn.f32 %f(39+i*5), %f(36+i*5), %f(38+i*5), %f(34+i*5);
ld.shared.f32 %f55, [%r4+24];
ld.shared.f32 %f56, [_ZZ24default_function_kernel0E20placeholder_d_shared+24];
fma.rn.f32 %f65, %f56, %f55, %f52;
ld.shared.f32 %f57, [%r4+80];
fma.rn.f32 %f66, %f56, %f57, %f54;
for.size.1 i in range(0, 2, 1):
add.s64 %rd(50+i*-1), %rd(50+i*-1), 28;
add.s32 %r90, %r90, 1;
setp.ne.s32 %p2, %r90, 112;
@%p2 bra $L__BB0_1;
cvta.to.global.u64 %rd45, %rd19;
shl.b32 %r88, %r5, 2;
add.s32 %r89, %r88, %r1;
mul.wide.s32 %rd46, %r89, 4;
add.s64 %rd47, %rd45, %rd46;
ld.global.nc.f32 %f58, [%rd47];
add.f32 %f59, %f65, %f58;
mov.f32 %f60, 0f00000000;
max.f32 %f61, %f59, %f60;
add.s64 %rd48, %rd2, %rd46;
st.global.f32 [%rd48], %f61;
ld.global.nc.f32 %f62, [%rd47+8];
add.f32 %f63, %f66, %f62;
max.f32 %f64, %f63, %f60;
st.global.f32 [%rd48+8], %f64;
ret;
}
| extern "C" __global__ void default_function_kernel0(float* __restrict__ placeholder, float* __restrict__ placeholder1, float* __restrict__ T_relu, float* __restrict__ placeholder2) {
float T_dense[2];
__shared__ float placeholder_d_shared[7];
__shared__ float placeholder_shared[28];
T_dense[0] = 0.000000e+00f;
T_dense[1] = 0.000000e+00f;
for (int k_outer_outer = 0; k_outer_outer < 112; ++k_outer_outer) {
__syncthreads();
int index0 = (int)threadIdx.x * 4;
int index1 = k_outer_outer * 7 + (int)threadIdx.x * 4;
placeholder_d_shared[index0] = placeholder[index1];
placeholder_d_shared[index0 + 1] = placeholder[index1 + 1];
placeholder_d_shared[index0 + 2] = placeholder[index1 + 2];
if ((int)threadIdx.x < 1) {
placeholder_d_shared[index0 + 3] = placeholder[index1 + 3];
}
int index2 = (int)threadIdx.x;
int index3 = (int)blockIdx.x * 3136 + k_outer_outer * 7 + (int)threadIdx.x;
placeholder_shared[index2] = placeholder1[index3];
placeholder_shared[index2 + 2] = placeholder1[index3 + 2];
placeholder_shared[index2 + 4] = placeholder1[index3 + 4];
int index4 = (int)blockIdx.x * 3136 + ((int)threadIdx.x + 6) / 7 * 784 + k_outer_outer * 7 + ((int)threadIdx.x + 6) % 7;
placeholder_shared[index2 + 6] = placeholder1[index4];
int index5 = (int)blockIdx.x * 3136 + ((int)threadIdx.x + 8) / 7 * 784 + k_outer_outer * 7 + ((int)threadIdx.x + 1);
placeholder_shared[index2 + 8] = placeholder1[index5];
int index6 = (int)blockIdx.x * 3136 + ((int)threadIdx.x + 10) / 7 * 784 + k_outer_outer * 7 + ((int)threadIdx.x + 3);
placeholder_shared[index2 + 10] = placeholder1[index6];
int index7 = (int)blockIdx.x * 3136 + ((int)threadIdx.x + 12) / 7 * 784 + k_outer_outer * 7 + ((int)threadIdx.x + 5);
placeholder_shared[index2 + 12] = placeholder1[index7];
placeholder_shared[index2 + 14] = placeholder1[index3 + 1568];
int index8 = (int)blockIdx.x * 3136 + ((int)threadIdx.x + 16) / 7 * 784 + k_outer_outer * 7 + ((int)threadIdx.x + 2);
placeholder_shared[index2 + 16] = placeholder1[index8];
int index9 = (int)blockIdx.x * 3136 + ((int)threadIdx.x + 18) / 7 * 784 + k_outer_outer * 7 + ((int)threadIdx.x + 4);
placeholder_shared[index2 + 18] = placeholder1[index9];
int index10 = (int)blockIdx.x * 3136 + ((int)threadIdx.x + 20) / 7 * 784 + k_outer_outer * 7 + ((int)threadIdx.x + 6) % 7;
placeholder_shared[index2 + 20] = placeholder1[index10];
int index11 = (int)blockIdx.x * 3136 + ((int)threadIdx.x + 22) / 7 * 784 + k_outer_outer * 7 + ((int)threadIdx.x + 1);
placeholder_shared[index2 + 22] = placeholder1[index11];
int index12 = (int)blockIdx.x * 3136 + ((int)threadIdx.x + 24) / 7 * 784 + k_outer_outer * 7 + ((int)threadIdx.x + 3);
placeholder_shared[index2 + 24] = placeholder1[index12];
int index13 = (int)blockIdx.x * 3136 + ((int)threadIdx.x + 26) / 7 * 784 + k_outer_outer * 7 + ((int)threadIdx.x + 5);
placeholder_shared[index2 + 26] = placeholder1[index13];
__syncthreads();
T_dense[0] = T_dense[0] + placeholder_d_shared[0] * placeholder_shared[index2 * 7];
int index14 = (int)threadIdx.x * 7 + 14;
T_dense[1] = T_dense[1] + placeholder_d_shared[0] * placeholder_shared[index14];
int index15 = (int)threadIdx.x * 7 + 1;
T_dense[0] = T_dense[0] + placeholder_d_shared[1] * placeholder_shared[index15];
int index16 = (int)threadIdx.x * 7 + 15;
T_dense[1] = T_dense[1] + placeholder_d_shared[1] * placeholder_shared[index16];
int index17 = (int)threadIdx.x * 7 + 2;
T_dense[0] = T_dense[0] + placeholder_d_shared[2] * placeholder_shared[index17];
int index18 = (int)threadIdx.x * 7 + 16;
T_dense[1] = T_dense[1] + placeholder_d_shared[2] * placeholder_shared[index18];
int index19 = (int)threadIdx.x * 7 + 3;
T_dense[0] = T_dense[0] + placeholder_d_shared[3] * placeholder_shared[index19];
int index20 = (int)threadIdx.x * 7 + 17;
T_dense[1] = T_dense[1] + placeholder_d_shared[3] * placeholder_shared[index20];
int index21 = (int)threadIdx.x * 7 + 4;
T_dense[0] = T_dense[0] + placeholder_d_shared[4] * placeholder_shared[index21];
int index22 = (int)threadIdx.x * 7 + 18;
T_dense[1] = T_dense[1] + placeholder_d_shared[4] * placeholder_shared[index22];
int index23 = (int)threadIdx.x * 7 + 5;
T_dense[0] = T_dense[0] + placeholder_d_shared[5] * placeholder_shared[index23];
int index24 = (int)threadIdx.x * 7 + 19;
T_dense[1] = T_dense[1] + placeholder_d_shared[5] * placeholder_shared[index24];
int index25 = (int)threadIdx.x * 7 + 6;
T_dense[0] = T_dense[0] + placeholder_d_shared[6] * placeholder_shared[index25];
int index26 = (int)threadIdx.x * 7 + 20;
T_dense[1] = T_dense[1] + placeholder_d_shared[6] * placeholder_shared[index26];
}
int index27 = (int)blockIdx.x * 4 + (int)threadIdx.x;
T_relu[index27] = max(T_dense[0] + placeholder2[index27], 0.000000e+00f);
T_relu[index27 + 2] = max(T_dense[1] + placeholder2[index27 + 2], 0.000000e+00f);
} |
[407f97a0b39db63cc31f3e635caadfd0,1,784,128,784,128,1,128] | tmpasbj8_ay | default_function_kernel0 | .visible .entry default_function_kernel0(
.param .u64 default_function_kernel0_param_0,
.param .u64 default_function_kernel0_param_1,
.param .u64 default_function_kernel0_param_2,
.param .u64 default_function_kernel0_param_3
)
{
.reg .pred %p<5>;
.reg .f32 %f<131>;
.reg .b32 %r<100>;
.reg .b64 %rd<22>;
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E20placeholder_d_shared[392];
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E18placeholder_shared[6272];
ld.param.u64 %rd6, [default_function_kernel0_param_0];
ld.param.u64 %rd7, [default_function_kernel0_param_1];
ld.param.u64 %rd8, [default_function_kernel0_param_2];
ld.param.u64 %rd9, [default_function_kernel0_param_3];
cvta.to.global.u64 %rd1, %rd7;
cvta.to.global.u64 %rd2, %rd6;
cvta.to.global.u64 %rd3, %rd8;
cvta.to.global.u64 %rd4, %rd9;
mov.u32 %r1, %tid.x;
shl.b32 %r27, %r1, 2;
mov.u32 %r28, _ZZ24default_function_kernel0E20placeholder_d_shared;
add.s32 %r2, %r28, %r27;
mov.u32 %r3, %ctaid.x;
for.size.1 i in range(0, 2, 1):
mul.lo.s32 %r(4+i*1), %r(3+i*-2), (12544+i*-12152);
shl.b32 %r6, %r1, 1;
add.s32 %r7, %r6, 24;
shl.b32 %r29, %r1, 3;
mov.u32 %r30, _ZZ24default_function_kernel0E18placeholder_shared;
add.s32 %r31, %r30, %r29;
add.s32 %r8, %r31, 64;
mov.u32 %r94, 0;
mov.f32 %f123, 0f00000000;
for.size.1 i in range(0, 3, 1):
mov.f32 %f(124+i*1), %f123;
$L__BB0_1:
bar.sync 0;
mul.lo.s32 %r10, %r94, 98;
add.s32 %r32, %r10, %r1;
mul.wide.s32 %rd10, %r32, 4;
add.s64 %rd5, %rd2, %rd10;
for.size.2 i in range(0, 24, 1):
ld.global.nc.f32 %f(21+i*1), [%rd5+(0+i*16)];
st.shared.f32 [%r2+(0+i*16)], %f(21+i*1);
setp.gt.s32 %p1, %r1, 1;
@%p1 bra $L__BB0_3;
ld.global.nc.f32 %f45, [%rd5+384];
st.shared.f32 [%r2+384], %f45;
$L__BB0_3:
add.s32 %r12, %r10, %r4;
mov.u32 %r98, 64;
mov.u32 %r95, %r8;
mov.u32 %r96, %r6;
mov.u32 %r97, %r7;
$L__BB0_4:
for.size.13 i in range(0, 2, 1):
mul.hi.s32 %r(34+i*13), %r(96+i*-50), 1402438301;
shr.u32 %r(35+i*13), %r(34+i*13), 31;
shr.s32 %r(36+i*13), %r(34+i*13), 5;
add.s32 %r(37+i*13), %r(36+i*13), %r(35+i*13);
mul.lo.s32 %r(38+i*13), %r(37+i*13), 98;
sub.s32 %r(39+i*13), %r(96+i*-50), %r(38+i*13);
add.s32 %r(40+i*13), %r12, %r(39+i*13);
mad.lo.s32 %r(41+i*13), %r(37+i*13), 784, %r(40+i*13);
mul.wide.s32 %rd(11+i*2), %r(41+i*13), 4;
add.s64 %rd(12+i*2), %rd1, %rd(11+i*2);
ld.global.nc.v2.u32 {%r(42+i*13), %r(43+i*13)}, [%rd(12+i*2)+0];
st.shared.v2.u32 [%r95+(-64+i*32)], {%r(42+i*13), %r(43+i*13)};
add.s32 %r(46+i*13), %r97, (-16+i*8);
for.size.12 i in range(0, 2, 1):
mul.hi.s32 %r(60+i*12), %r(59+i*38), 1402438301;
shr.u32 %r(61+i*12), %r(60+i*12), 31;
shr.s32 %r(62+i*12), %r(60+i*12), 5;
add.s32 %r(63+i*12), %r(62+i*12), %r(61+i*12);
mul.lo.s32 %r(64+i*12), %r(63+i*12), 98;
sub.s32 %r(65+i*12), %r(59+i*38), %r(64+i*12);
add.s32 %r(66+i*12), %r12, %r(65+i*12);
mad.lo.s32 %r(67+i*12), %r(63+i*12), 784, %r(66+i*12);
mul.wide.s32 %rd(15+i*2), %r(67+i*12), 4;
add.s64 %rd(16+i*2), %rd1, %rd(15+i*2);
ld.global.nc.v2.u32 {%r(68+i*12), %r(69+i*12)}, [%rd(16+i*2)+0];
st.shared.v2.u32 [%r95+(0+i*32)], {%r(68+i*12), %r(69+i*12)};
for.size.1 i in range(0, 2, 1):
add.s32 %r(97+i*-1), %r(97+i*-1), 32;
for.size.1 i in range(0, 2, 1):
add.s32 %r(95+i*3), %r(95+i*3), 128;
setp.ne.s32 %p2, %r98, 6336;
@%p2 bra $L__BB0_4;
bar.sync 0;
mov.u32 %r99, 0;
bra.uni $L__BB0_6;
$L__BB0_7:
for.size.1 i in range(0, 2, 1):
ld.shared.f32 %f(60+i*1), [%r(23+i*-1)+8];
for.size.2 i in range(0, 2, 1):
fma.rn.f32 %f(62+i*2), %f61, %f(60+i*3), %f(126+i*-1);
ld.shared.f32 %f(63+i*2), [%r23+(400+i*392)];
for.size.2 i in range(0, 2, 1):
fma.rn.f32 %f(66+i*2), %f61, %f(65+i*2), %f(124+i*-1);
ld.shared.f32 %f(67+i*2), [%r23+(1184+i*-1172)];
for.size.9 i in range(0, 4, 1):
ld.shared.f32 %f(70+i*9), [%r22+(12+i*4)];
fma.rn.f32 %f(71+i*9), %f(70+i*9), %f(69+i*9), %f(62+i*9);
ld.shared.f32 %f(72+i*9), [%r23+(404+i*4)];
fma.rn.f32 %f(73+i*9), %f(70+i*9), %f(72+i*9), %f(64+i*9);
ld.shared.f32 %f(74+i*9), [%r23+(796+i*4)];
fma.rn.f32 %f(75+i*9), %f(70+i*9), %f(74+i*9), %f(66+i*9);
ld.shared.f32 %f(76+i*9), [%r23+(1188+i*4)];
fma.rn.f32 %f(77+i*9), %f(70+i*9), %f(76+i*9), %f(68+i*9);
ld.shared.f32 %f(78+i*9), [%r23+(16+i*4)];
ld.shared.f32 %f106, [%r22+28];
for.size.2 i in range(0, 2, 1):
fma.rn.f32 %f(126+i*-1), %f106, %f(105+i*2), %f(98+i*2);
ld.shared.f32 %f(107+i*1), [%r23+(420+i*392)];
fma.rn.f32 %f124, %f106, %f108, %f102;
ld.shared.f32 %f109, [%r23+1204];
fma.rn.f32 %f123, %f106, %f109, %f104;
add.s32 %r99, %r99, 4;
$L__BB0_6:
for.size.1 i in range(0, 2, 1):
shl.b32 %r(85+i*1), %r99, (1+i*2);
for.size.1 i in range(0, 2, 1):
add.s32 %r(22+i*66), %r(28+i*57), %r(86+i*-81);
shl.b32 %r89, %r88, 2;
add.s32 %r23, %r30, %r89;
for.size.1 i in range(0, 2, 1):
ld.shared.f32 %f(46+i*1), [%r(23+i*-1)+0];
for.size.2 i in range(0, 2, 1):
fma.rn.f32 %f(48+i*2), %f47, %f(46+i*3), %f(126+i*-1);
ld.shared.f32 %f(49+i*2), [%r23+(392+i*392)];
for.size.2 i in range(0, 2, 1):
fma.rn.f32 %f(52+i*2), %f47, %f(51+i*2), %f(124+i*-1);
ld.shared.f32 %f(53+i*2), [%r23+(1176+i*-1172)];
ld.shared.f32 %f56, [%r22+4];
for.size.2 i in range(0, 2, 1):
fma.rn.f32 %f(126+i*-1), %f56, %f(55+i*2), %f(48+i*2);
ld.shared.f32 %f(57+i*1), [%r23+(396+i*392)];
fma.rn.f32 %f124, %f56, %f58, %f52;
ld.shared.f32 %f59, [%r23+1180];
fma.rn.f32 %f123, %f56, %f59, %f54;
setp.eq.s32 %p3, %r99, 48;
@%p3 bra $L__BB0_8;
bra.uni $L__BB0_7;
$L__BB0_8:
add.s32 %r94, %r94, 1;
setp.lt.u32 %p4, %r94, 8;
@%p4 bra $L__BB0_1;
shl.b32 %r91, %r3, 4;
add.s32 %r93, %r91, %r27;
mul.wide.s32 %rd19, %r93, 4;
add.s64 %rd20, %rd4, %rd19;
ld.global.nc.f32 %f110, [%rd20];
add.f32 %f111, %f126, %f110;
mov.f32 %f112, 0f00000000;
max.f32 %f113, %f111, %f112;
add.s64 %rd21, %rd3, %rd19;
for.size.4 i in range(0, 3, 1):
st.global.f32 [%rd21+(0+i*4)], %f(113+i*3);
ld.global.nc.f32 %f(114+i*3), [%rd20+(4+i*4)];
add.f32 %f(115+i*3), %f(125+i*-1), %f(114+i*3);
max.f32 %f(116+i*3), %f(115+i*3), %f112;
st.global.f32 [%rd21+12], %f122;
ret;
}
| extern "C" __global__ void default_function_kernel0(float* __restrict__ placeholder, float* __restrict__ placeholder1, float* __restrict__ T_relu, float* __restrict__ placeholder2) {
float T_dense[4];
__shared__ float placeholder_d_shared[98];
__shared__ float placeholder_shared[1568];
T_dense[0] = 0.000000e+00f;
T_dense[1] = 0.000000e+00f;
T_dense[2] = 0.000000e+00f;
T_dense[3] = 0.000000e+00f;
for (int k_outer_outer = 0; k_outer_outer < 8; ++k_outer_outer) {
__syncthreads();
int index0 = (int)threadIdx.x;
int index1 = k_outer_outer * 98 + (int)threadIdx.x;
placeholder_d_shared[index0] = placeholder[index1];
placeholder_d_shared[index0 + 4] = placeholder[index1 + 4];
placeholder_d_shared[index0 + 8] = placeholder[index1 + 8];
placeholder_d_shared[index0 + 12] = placeholder[index1 + 12];
placeholder_d_shared[index0 + 16] = placeholder[index1 + 16];
placeholder_d_shared[index0 + 20] = placeholder[index1 + 20];
placeholder_d_shared[index0 + 24] = placeholder[index1 + 24];
placeholder_d_shared[index0 + 28] = placeholder[index1 + 28];
placeholder_d_shared[index0 + 32] = placeholder[index1 + 32];
placeholder_d_shared[index0 + 36] = placeholder[index1 + 36];
placeholder_d_shared[index0 + 40] = placeholder[index1 + 40];
placeholder_d_shared[index0 + 44] = placeholder[index1 + 44];
placeholder_d_shared[index0 + 48] = placeholder[index1 + 48];
placeholder_d_shared[index0 + 52] = placeholder[index1 + 52];
placeholder_d_shared[index0 + 56] = placeholder[index1 + 56];
placeholder_d_shared[index0 + 60] = placeholder[index1 + 60];
placeholder_d_shared[index0 + 64] = placeholder[index1 + 64];
placeholder_d_shared[index0 + 68] = placeholder[index1 + 68];
placeholder_d_shared[index0 + 72] = placeholder[index1 + 72];
placeholder_d_shared[index0 + 76] = placeholder[index1 + 76];
placeholder_d_shared[index0 + 80] = placeholder[index1 + 80];
placeholder_d_shared[index0 + 84] = placeholder[index1 + 84];
placeholder_d_shared[index0 + 88] = placeholder[index1 + 88];
placeholder_d_shared[index0 + 92] = placeholder[index1 + 92];
if ((int)threadIdx.x < 2) {
placeholder_d_shared[index0 + 96] = placeholder[index1 + 96];
}
for (int ax0_ax1_fused_outer_outer = 0; ax0_ax1_fused_outer_outer < 196; ++ax0_ax1_fused_outer_outer) {
((float2*)(placeholder_shared + ax0_ax1_fused_outer_outer * 8 + (int)threadIdx.x * 2))[0] = ((float2*)(placeholder1 + (int)blockIdx.x * 12544 + (ax0_ax1_fused_outer_outer * 8 + (int)threadIdx.x * 2) / 98 * 784 + k_outer_outer * 98 + (ax0_ax1_fused_outer_outer * 8 + (int)threadIdx.x * 2) % 98))[0];
}
__syncthreads();
for (int k_outer_inner = 0; k_outer_inner < 49; ++k_outer_inner) {
int index2 = k_outer_inner * 2;
int index3 = (int)threadIdx.x * 392 + k_outer_inner * 2;
T_dense[0] = T_dense[0] + placeholder_d_shared[index2] * placeholder_shared[index3];
T_dense[1] = T_dense[1] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 98]);
T_dense[2] = T_dense[2] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 196]);
T_dense[3] = T_dense[3] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 294]);
T_dense[0] = T_dense[0] + (placeholder_d_shared[index2 + 1] * placeholder_shared[index3 + 1]);
T_dense[1] = T_dense[1] + (placeholder_d_shared[index2 + 1] * placeholder_shared[index3 + 99]);
T_dense[2] = T_dense[2] + (placeholder_d_shared[index2 + 1] * placeholder_shared[index3 + 197]);
T_dense[3] = T_dense[3] + (placeholder_d_shared[index2 + 1] * placeholder_shared[index3 + 295]);
}
}
for (int ax1_inner = 0; ax1_inner < 4; ++ax1_inner) {
int index4 = (int)blockIdx.x * 16 + (int)threadIdx.x * 4 + ax1_inner;
int index5 = ax1_inner;
T_relu[index4] = max(T_dense[index5] + placeholder2[index4], 0.000000e+00f);
}
} |
[407f97a0b39db63cc31f3e635caadfd0,1,784,128,784,128,1,128] | tmpc6ax8elt | default_function_kernel0 | .visible .entry default_function_kernel0(
.param .u64 default_function_kernel0_param_0,
.param .u64 default_function_kernel0_param_1,
.param .u64 default_function_kernel0_param_2,
.param .u64 default_function_kernel0_param_3
)
{
.reg .pred %p<5>;
.reg .f32 %f<71>;
.reg .b32 %r<97>;
.reg .b64 %rd<18>;
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E20placeholder_d_shared[1568];
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E18placeholder_shared[100352];
ld.param.u64 %rd6, [default_function_kernel0_param_0];
ld.param.u64 %rd7, [default_function_kernel0_param_1];
ld.param.u64 %rd4, [default_function_kernel0_param_2];
ld.param.u64 %rd5, [default_function_kernel0_param_3];
cvta.to.global.u64 %rd1, %rd7;
cvta.to.global.u64 %rd2, %rd6;
mov.u32 %r1, %tid.x;
shl.b32 %r21, %r1, 2;
mov.u32 %r22, _ZZ24default_function_kernel0E20placeholder_d_shared;
add.s32 %r2, %r22, %r21;
mov.u32 %r23, %ctaid.x;
for.size.1 i in range(0, 2, 1):
mul.lo.s32 %r(3+i*1), %r(23+i*-22), (50176+i*-49392);
shl.b32 %r24, %r1, 3;
add.s32 %r5, %r24, 256;
shl.b32 %r25, %r1, 5;
mov.u32 %r26, _ZZ24default_function_kernel0E18placeholder_shared;
add.s32 %r27, %r26, %r25;
add.s32 %r6, %r27, 1040;
mov.u32 %r92, 0;
mov.f32 %f67, 0f00000000;
mov.f32 %f68, %f67;
$L__BB0_1:
bar.sync 0;
mul.lo.s32 %r8, %r92, 392;
add.s32 %r28, %r8, %r1;
mul.wide.s32 %rd8, %r28, 4;
add.s64 %rd3, %rd2, %rd8;
for.size.2 i in range(0, 12, 1):
ld.global.nc.f32 %f(9+i*1), [%rd3+(0+i*128)];
st.shared.f32 [%r2+(0+i*128)], %f(9+i*1);
setp.gt.s32 %p1, %r1, 7;
@%p1 bra $L__BB0_3;
ld.global.nc.f32 %f21, [%rd3+1536];
st.shared.f32 [%r2+1536], %f21;
$L__BB0_3:
add.s32 %r10, %r8, %r3;
mov.u32 %r95, 1024;
mov.u32 %r93, %r6;
mov.u32 %r94, %r5;
$L__BB0_4:
add.s32 %r30, %r94, -256;
for.size.14 i in range(0, 2, 1):
mul.hi.s32 %r(31+i*24), %r(30+i*64), 1402438301;
shr.u32 %r(32+i*24), %r(31+i*24), 31;
shr.s32 %r(33+i*24), %r(31+i*24), 7;
add.s32 %r(34+i*24), %r(33+i*24), %r(32+i*24);
mul.lo.s32 %r(35+i*24), %r(34+i*24), 392;
sub.s32 %r(36+i*24), %r(30+i*64), %r(35+i*24);
add.s32 %r(37+i*24), %r10, %r(36+i*24);
mad.lo.s32 %r(38+i*24), %r(34+i*24), 784, %r(37+i*24);
mul.wide.s32 %rd(9+i*2), %r(38+i*24), 4;
add.s64 %rd(10+i*2), %rd1, %rd(9+i*2);
ld.global.nc.v4.u32 {%r(39+i*24), %r(40+i*24), %r(41+i*24), %r(42+i*24)}, [%rd(10+i*2)+0];
ld.global.nc.v4.u32 {%r(47+i*24), %r(48+i*24), %r(49+i*24), %r(50+i*24)}, [%rd(10+i*2)+16];
st.shared.v4.u32 [%r93+(-1040+i*1024)], {%r(39+i*24), %r(40+i*24), %r(41+i*24), %r(42+i*24)};
st.shared.v4.u32 [%r93+(-1024+i*1024)], {%r(47+i*24), %r(48+i*24), %r(49+i*24), %r(50+i*24)};
for.size.1 i in range(0, 2, 1):
add.s32 %r(94+i*-1), %r(94+i*-1), (512+i*1536);
add.s32 %r95, %r95, 2048;
setp.ne.s32 %p2, %r95, 101376;
@%p2 bra $L__BB0_4;
bar.sync 0;
mov.u32 %r96, 0;
$L__BB0_6:
for.size.1 i in range(0, 2, 1):
shl.b32 %r(80+i*1), %r96, (1+i*2);
for.size.1 i in range(0, 2, 1):
add.s32 %r(83+i*1), %r(22+i*58), %r(81+i*-77);
shl.b32 %r85, %r84, 2;
add.s32 %r87, %r26, %r85;
for.size.1 i in range(0, 2, 1):
ld.shared.f32 %f(22+i*1), [%r(87+i*-4)+0];
for.size.10 i in range(0, 2, 1):
fma.rn.f32 %f(24+i*10), %f(23+i*10), %f(22+i*10), %f(68+i*-41);
ld.shared.f32 %f(25+i*10), [%r87+(4+i*8)];
ld.shared.f32 %f(26+i*10), [%r83+(4+i*8)];
fma.rn.f32 %f(27+i*10), %f(26+i*10), %f(25+i*10), %f(24+i*10);
ld.shared.f32 %f(28+i*10), [%r87+(1568+i*8)];
fma.rn.f32 %f(29+i*10), %f(23+i*10), %f(28+i*10), %f(67+i*-36);
ld.shared.f32 %f(30+i*10), [%r87+(1572+i*8)];
fma.rn.f32 %f(31+i*10), %f(26+i*10), %f(30+i*10), %f(29+i*10);
ld.shared.f32 %f(32+i*10), [%r87+(8+i*8)];
ld.shared.f32 %f(33+i*10), [%r83+(8+i*8)];
fma.rn.f32 %f44, %f43, %f42, %f37;
for.size.1 i in range(0, 2, 1):
ld.shared.f32 %f(45+i*1), [%r(87+i*-4)+20];
for.size.2 i in range(0, 2, 1):
fma.rn.f32 %f(47+i*2), %f(46+i*-3), %f(45+i*3), %f(44+i*-3);
ld.shared.f32 %f(48+i*2), [%r87+(1584+i*4)];
for.size.3 i in range(0, 2, 1):
fma.rn.f32 %f(51+i*3), %f(46+i*7), %f(50+i*2), %f(49+i*-2);
ld.shared.f32 %f(52+i*3), [%r87+(24+i*4)];
ld.shared.f32 %f(53+i*3), [%r83+(24+i*4)];
for.size.2 i in range(0, 2, 1):
fma.rn.f32 %f(68+i*-10), %f(56+i*-3), %f(55+i*2), %f(54+i*-3);
ld.shared.f32 %f(57+i*2), [%r87+(1592+i*4)];
fma.rn.f32 %f67, %f56, %f59, %f58;
add.s32 %r96, %r96, 4;
setp.ne.s32 %p3, %r96, 196;
@%p3 bra $L__BB0_6;
add.s32 %r92, %r92, 1;
setp.lt.u32 %p4, %r92, 2;
@%p4 bra $L__BB0_1;
for.size.1 i in range(0, 2, 1):
shl.b32 %r(89+i*1), %r(23+i*-22), (6+i*-5);
add.s32 %r91, %r89, %r90;
cvta.to.global.u64 %rd13, %rd5;
mul.wide.s32 %rd14, %r91, 4;
add.s64 %rd15, %rd13, %rd14;
ld.global.nc.f32 %f60, [%rd15];
add.f32 %f61, %f68, %f60;
mov.f32 %f62, 0f00000000;
max.f32 %f63, %f61, %f62;
cvta.to.global.u64 %rd16, %rd4;
add.s64 %rd17, %rd16, %rd14;
st.global.f32 [%rd17], %f63;
ld.global.nc.f32 %f64, [%rd15+4];
add.f32 %f65, %f67, %f64;
max.f32 %f66, %f65, %f62;
st.global.f32 [%rd17+4], %f66;
ret;
}
| extern "C" __global__ void default_function_kernel0(float* __restrict__ placeholder, float* __restrict__ placeholder1, float* __restrict__ T_relu, float* __restrict__ placeholder2) {
float T_dense[2];
__shared__ float placeholder_d_shared[392];
__shared__ float placeholder_shared[25088];
T_dense[0] = 0.000000e+00f;
T_dense[1] = 0.000000e+00f;
for (int k_outer_outer = 0; k_outer_outer < 2; ++k_outer_outer) {
__syncthreads();
int index0 = (int)threadIdx.x;
int index1 = k_outer_outer * 392 + (int)threadIdx.x;
placeholder_d_shared[index0] = placeholder[index1];
placeholder_d_shared[index0 + 32] = placeholder[index1 + 32];
placeholder_d_shared[index0 + 64] = placeholder[index1 + 64];
placeholder_d_shared[index0 + 96] = placeholder[index1 + 96];
placeholder_d_shared[index0 + 128] = placeholder[index1 + 128];
placeholder_d_shared[index0 + 160] = placeholder[index1 + 160];
placeholder_d_shared[index0 + 192] = placeholder[index1 + 192];
placeholder_d_shared[index0 + 224] = placeholder[index1 + 224];
placeholder_d_shared[index0 + 256] = placeholder[index1 + 256];
placeholder_d_shared[index0 + 288] = placeholder[index1 + 288];
placeholder_d_shared[index0 + 320] = placeholder[index1 + 320];
placeholder_d_shared[index0 + 352] = placeholder[index1 + 352];
if ((int)threadIdx.x < 8) {
placeholder_d_shared[index0 + 384] = placeholder[index1 + 384];
}
for (int ax0_ax1_fused_outer_outer = 0; ax0_ax1_fused_outer_outer < 98; ++ax0_ax1_fused_outer_outer) {
ulonglong4*(placeholder_shared + ax0_ax1_fused_outer_outer * 256 + (int)threadIdx.x * 8)[0] = ulonglong4*(placeholder1 + (int)blockIdx.x * 50176 + (ax0_ax1_fused_outer_outer * 256 + (int)threadIdx.x * 8) / 392 * 784 + k_outer_outer * 392 + (ax0_ax1_fused_outer_outer * 256 + (int)threadIdx.x * 8) % 392)[0];
}
__syncthreads();
for (int k_outer_inner = 0; k_outer_inner < 196; ++k_outer_inner) {
int index2 = k_outer_inner * 2;
int index3 = (int)threadIdx.x * 784 + k_outer_inner * 2;
T_dense[0] = T_dense[0] + placeholder_d_shared[index2] * placeholder_shared[index3];
T_dense[0] = T_dense[0] + (placeholder_d_shared[index2 + 1] * placeholder_shared[index3 + 1]);
T_dense[1] = T_dense[1] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 392]);
T_dense[1] = T_dense[1] + (placeholder_d_shared[index2 + 1] * placeholder_shared[index3 + 393]);
}
}
for (int ax1_inner = 0; ax1_inner < 2; ++ax1_inner) {
int index4 = (int)blockIdx.x * 64 + (int)threadIdx.x * 2 + ax1_inner;
int index5 = ax1_inner;
T_relu[index4] = max(T_dense[index5] + placeholder2[index4], 0.000000e+00f);
}
} |
[407f97a0b39db63cc31f3e635caadfd0,1,784,128,784,128,1,128] | tmpjaj0lm38 | default_function_kernel0 | .visible .entry default_function_kernel0(
.param .u64 default_function_kernel0_param_0,
.param .u64 default_function_kernel0_param_1,
.param .u64 default_function_kernel0_param_2,
.param .u64 default_function_kernel0_param_3
)
{
.reg .pred %p<5>;
.reg .f32 %f<651>;
.reg .b32 %r<128>;
.reg .b64 %rd<38>;
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E20placeholder_d_shared[1568];
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E18placeholder_shared[200704];
ld.param.u64 %rd8, [default_function_kernel0_param_0];
ld.param.u64 %rd9, [default_function_kernel0_param_1];
cvta.to.global.u64 %rd12, %rd8;
cvta.to.global.u64 %rd2, %rd9;
mov.u32 %r1, %tid.x;
mul.lo.s32 %r2, %r1, 25088;
shl.b32 %r17, %r1, 3;
mov.u32 %r18, _ZZ24default_function_kernel0E20placeholder_d_shared;
add.s32 %r4, %r18, %r17;
add.s64 %rd4, %rd12, 16;
mov.u32 %r123, 0;
mov.f32 %f523, 0f00000000;
for.size.1 i in range(0, 63, 1):
mov.f32 %f(524+i*1), %f523;
$L__BB0_1:
mov.u32 %r124, 0;
mov.u32 %r121, %tid.x;
shl.b32 %r120, %r121, 1;
bar.sync 0;
mul.lo.s32 %r6, %r123, 392;
add.s32 %r21, %r120, %r6;
mul.wide.s32 %rd13, %r21, 4;
add.s64 %rd37, %rd4, %rd13;
mov.u32 %r125, 16;
$L__BB0_2:
ld.global.nc.v2.u32 {%r22, %r23}, [%rd37+-16];
add.s32 %r26, %r4, %r124;
st.shared.v2.u32 [%r26], {%r22, %r23};
ld.global.nc.v2.u32 {%r27, %r28}, [%rd37];
st.shared.v2.u32 [%r26+16], {%r27, %r28};
add.s64 %rd37, %rd37, 32;
add.s32 %r9, %r125, 32;
setp.ne.s32 %p1, %r9, 1584;
add.s32 %r124, %r125, 16;
mov.u32 %r125, %r9;
@%p1 bra $L__BB0_2;
mov.u32 %r126, 0;
$L__BB0_4:
mul.lo.s32 %r117, %r123, 392;
shl.b32 %r32, %r126, 1;
add.s32 %r33, %r32, %r1;
mul.hi.s32 %r34, %r33, 1402438301;
shr.u32 %r35, %r34, 31;
shr.s32 %r36, %r34, 7;
add.s32 %r37, %r36, %r35;
mul.lo.s32 %r38, %r37, 392;
sub.s32 %r39, %r33, %r38;
add.s32 %r40, %r39, %r117;
mad.lo.s32 %r41, %r37, 784, %r40;
mul.wide.s32 %rd14, %r41, 4;
add.s64 %rd15, %rd2, %rd14;
ld.global.nc.f32 %f257, [%rd15];
shl.b32 %r42, %r33, 2;
mov.u32 %r43, _ZZ24default_function_kernel0E18placeholder_shared;
add.s32 %r44, %r43, %r42;
for.size.13 i in range(0, 7, 1):
st.shared.f32 [%r44+(0+i*8)], %f(257+i*1);
add.s32 %r(45+i*9), %r33, (2+i*2);
mul.hi.s32 %r(46+i*9), %r(45+i*9), 1402438301;
shr.u32 %r(47+i*9), %r(46+i*9), 31;
shr.s32 %r(48+i*9), %r(46+i*9), 7;
add.s32 %r(49+i*9), %r(48+i*9), %r(47+i*9);
mul.lo.s32 %r(50+i*9), %r(49+i*9), 392;
sub.s32 %r(51+i*9), %r(45+i*9), %r(50+i*9);
add.s32 %r(52+i*9), %r(51+i*9), %r117;
mad.lo.s32 %r(53+i*9), %r(49+i*9), 784, %r(52+i*9);
mul.wide.s32 %rd(16+i*2), %r(53+i*9), 4;
add.s64 %rd(17+i*2), %rd2, %rd(16+i*2);
ld.global.nc.f32 %f(258+i*1), [%rd(17+i*2)+0];
st.shared.f32 [%r44+56], %f264;
add.s32 %r126, %r126, 8;
setp.ne.s32 %p2, %r126, 25088;
@%p2 bra $L__BB0_4;
bar.sync 0;
mov.u32 %r127, 0;
$L__BB0_6:
mov.u32 %r119, _ZZ24default_function_kernel0E20placeholder_d_shared;
mov.u32 %r118, _ZZ24default_function_kernel0E18placeholder_shared;
add.s32 %r109, %r127, %r2;
for.size.3 i in range(0, 2, 1):
shl.b32 %r(110+i*3), %r(109+i*18), 2;
add.s32 %r(112+i*3), %r(118+i*1), %r(110+i*3);
ld.shared.f32 %f(265+i*1), [%r(112+i*3)+0];
for.size.2 i in range(0, 2, 1):
fma.rn.f32 %f(586+i*-1), %f266, %f(265+i*2), %f(586+i*-1);
ld.shared.f32 %f(267+i*1), [%r112+(1568+i*1568)];
for.size.2 i in range(0, 61, 1):
fma.rn.f32 %f(584+i*-1), %f266, %f(268+i*1), %f(584+i*-1);
ld.shared.f32 %f(269+i*1), [%r112+(4704+i*1568)];
fma.rn.f32 %f523, %f266, %f329, %f523;
add.s32 %r127, %r127, 1;
setp.ne.s32 %p3, %r127, 392;
@%p3 bra $L__BB0_6;
add.s32 %r123, %r123, 1;
setp.lt.u32 %p4, %r123, 2;
@%p4 bra $L__BB0_1;
ld.param.u64 %rd36, [default_function_kernel0_param_2];
cvta.to.global.u64 %rd35, %rd36;
ld.param.u64 %rd34, [default_function_kernel0_param_3];
cvta.to.global.u64 %rd33, %rd34;
shl.b32 %r116, %r1, 6;
mul.wide.s32 %rd30, %r116, 4;
add.s64 %rd31, %rd33, %rd30;
ld.global.nc.f32 %f330, [%rd31];
add.f32 %f331, %f586, %f330;
mov.f32 %f332, 0f00000000;
max.f32 %f333, %f331, %f332;
add.s64 %rd32, %rd35, %rd30;
for.size.4 i in range(0, 63, 1):
st.global.f32 [%rd32+(0+i*4)], %f(333+i*3);
ld.global.nc.f32 %f(334+i*3), [%rd31+(4+i*4)];
add.f32 %f(335+i*3), %f(585+i*-1), %f(334+i*3);
max.f32 %f(336+i*3), %f(335+i*3), %f332;
st.global.f32 [%rd32+252], %f522;
ret;
}
| extern "C" __global__ void default_function_kernel0(float* __restrict__ placeholder, float* __restrict__ placeholder1, float* __restrict__ T_relu, float* __restrict__ placeholder2) {
float T_dense[64];
__shared__ float placeholder_d_shared[392];
__shared__ float placeholder_shared[50176];
for (int j_inner_init = 0; j_inner_init < 64; ++j_inner_init) {
int index0 = j_inner_init;
T_dense[index0] = 0.000000e+00f;
}
for (int k_outer_outer = 0; k_outer_outer < 2; ++k_outer_outer) {
__syncthreads();
for (int ax0_ax1_fused_outer_outer = 0; ax0_ax1_fused_outer_outer < 98; ++ax0_ax1_fused_outer_outer) {
((float2*)(placeholder_d_shared + ax0_ax1_fused_outer_outer * 4 + (int)threadIdx.x * 2))[0] = ((float2*)(placeholder + k_outer_outer * 392 + ax0_ax1_fused_outer_outer * 4 + (int)threadIdx.x * 2))[0];
}
for (int ax0_ax1_fused_outer_outer1 = 0; ax0_ax1_fused_outer_outer1 < 25088; ++ax0_ax1_fused_outer_outer1) {
int index1 = ax0_ax1_fused_outer_outer1 * 2 + (int)threadIdx.x;
int index2 = (ax0_ax1_fused_outer_outer1 * 2 + (int)threadIdx.x) / 392 * 784 + k_outer_outer * 392 + (ax0_ax1_fused_outer_outer1 * 2 + (int)threadIdx.x) % 392;
placeholder_shared[index1] = placeholder1[index2];
}
__syncthreads();
for (int k_outer_inner = 0; k_outer_inner < 392; ++k_outer_inner) {
for (int j_inner = 0; j_inner < 64; ++j_inner) {
int index3 = j_inner;
int index4 = k_outer_inner;
int index5 = (int)threadIdx.x * 25088 + j_inner * 392 + k_outer_inner;
T_dense[index3] = T_dense[index3] + placeholder_d_shared[index4] * placeholder_shared[index5];
}
}
}
for (int ax1_inner = 0; ax1_inner < 64; ++ax1_inner) {
int index6 = (int)threadIdx.x * 64 + ax1_inner;
int index7 = ax1_inner;
T_relu[index6] = max(T_dense[index7] + placeholder2[index6], 0.000000e+00f);
}
} |
[407f97a0b39db63cc31f3e635caadfd0,1,784,128,784,128,1,128] | tmpzybl8g4k | default_function_kernel0 | .visible .entry default_function_kernel0(
.param .u64 default_function_kernel0_param_0,
.param .u64 default_function_kernel0_param_1,
.param .u64 default_function_kernel0_param_2,
.param .u64 default_function_kernel0_param_3
)
{
.reg .pred %p<87>;
.reg .f32 %f<82>;
.reg .b32 %r<94>;
.reg .b64 %rd<20>;
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E20placeholder_d_shared[3136];
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E18placeholder_shared[401408];
ld.param.u64 %rd10, [default_function_kernel0_param_0];
ld.param.u64 %rd7, [default_function_kernel0_param_1];
ld.param.u64 %rd8, [default_function_kernel0_param_2];
ld.param.u64 %rd9, [default_function_kernel0_param_3];
cvta.to.global.u64 %rd11, %rd10;
mov.u32 %r1, %tid.x;
mul.lo.s32 %r2, %r1, 28;
for.size.1 i in range(0, 2, 1):
setp.gt.s32 %p(1+i*1), %r(1+i*1), (27+i*756);
mul.wide.s32 %rd12, %r2, 4;
add.s64 %rd1, %rd11, %rd12;
shl.b32 %r14, %r2, 2;
mov.u32 %r15, _ZZ24default_function_kernel0E20placeholder_d_shared;
add.s32 %r3, %r15, %r14;
or.pred %p3, %p1, %p2;
@%p3 bra $L__BB0_2;
ld.global.nc.f32 %f5, [%rd1];
st.shared.f32 [%r3], %f5;
$L__BB0_2:
setp.gt.s32 %p5, %r2, 782;
or.pred %p6, %p1, %p5;
@%p6 bra $L__BB0_4;
ld.global.nc.f32 %f6, [%rd1+4];
st.shared.f32 [%r3+4], %f6;
$L__BB0_4:
setp.gt.s32 %p8, %r2, 781;
or.pred %p9, %p1, %p8;
@%p9 bra $L__BB0_6;
ld.global.nc.f32 %f7, [%rd1+8];
st.shared.f32 [%r3+8], %f7;
$L__BB0_6:
setp.gt.s32 %p11, %r2, 780;
or.pred %p12, %p1, %p11;
@%p12 bra $L__BB0_8;
ld.global.nc.f32 %f8, [%rd1+12];
st.shared.f32 [%r3+12], %f8;
$L__BB0_8:
setp.gt.s32 %p14, %r2, 779;
or.pred %p15, %p1, %p14;
@%p15 bra $L__BB0_10;
ld.global.nc.f32 %f9, [%rd1+16];
st.shared.f32 [%r3+16], %f9;
$L__BB0_10:
setp.gt.s32 %p17, %r2, 778;
or.pred %p18, %p1, %p17;
@%p18 bra $L__BB0_12;
ld.global.nc.f32 %f10, [%rd1+20];
st.shared.f32 [%r3+20], %f10;
$L__BB0_12:
setp.gt.s32 %p20, %r2, 777;
or.pred %p21, %p1, %p20;
@%p21 bra $L__BB0_14;
ld.global.nc.f32 %f11, [%rd1+24];
st.shared.f32 [%r3+24], %f11;
$L__BB0_14:
setp.gt.s32 %p23, %r2, 776;
or.pred %p24, %p1, %p23;
@%p24 bra $L__BB0_16;
ld.global.nc.f32 %f12, [%rd1+28];
st.shared.f32 [%r3+28], %f12;
$L__BB0_16:
setp.gt.s32 %p26, %r2, 775;
or.pred %p27, %p1, %p26;
@%p27 bra $L__BB0_18;
ld.global.nc.f32 %f13, [%rd1+32];
st.shared.f32 [%r3+32], %f13;
$L__BB0_18:
setp.gt.s32 %p29, %r2, 774;
or.pred %p30, %p1, %p29;
@%p30 bra $L__BB0_20;
ld.global.nc.f32 %f14, [%rd1+36];
st.shared.f32 [%r3+36], %f14;
$L__BB0_20:
setp.gt.s32 %p32, %r2, 773;
or.pred %p33, %p1, %p32;
@%p33 bra $L__BB0_22;
ld.global.nc.f32 %f15, [%rd1+40];
st.shared.f32 [%r3+40], %f15;
$L__BB0_22:
setp.gt.s32 %p35, %r2, 772;
or.pred %p36, %p1, %p35;
@%p36 bra $L__BB0_24;
ld.global.nc.f32 %f16, [%rd1+44];
st.shared.f32 [%r3+44], %f16;
$L__BB0_24:
setp.gt.s32 %p38, %r2, 771;
or.pred %p39, %p1, %p38;
@%p39 bra $L__BB0_26;
ld.global.nc.f32 %f17, [%rd1+48];
st.shared.f32 [%r3+48], %f17;
$L__BB0_26:
setp.gt.s32 %p41, %r2, 770;
or.pred %p42, %p1, %p41;
@%p42 bra $L__BB0_28;
ld.global.nc.f32 %f18, [%rd1+52];
st.shared.f32 [%r3+52], %f18;
$L__BB0_28:
setp.gt.s32 %p44, %r2, 769;
or.pred %p45, %p1, %p44;
@%p45 bra $L__BB0_30;
ld.global.nc.f32 %f19, [%rd1+56];
st.shared.f32 [%r3+56], %f19;
$L__BB0_30:
setp.gt.s32 %p47, %r2, 768;
or.pred %p48, %p1, %p47;
@%p48 bra $L__BB0_32;
ld.global.nc.f32 %f20, [%rd1+60];
st.shared.f32 [%r3+60], %f20;
$L__BB0_32:
setp.gt.s32 %p50, %r2, 767;
or.pred %p51, %p1, %p50;
@%p51 bra $L__BB0_34;
ld.global.nc.f32 %f21, [%rd1+64];
st.shared.f32 [%r3+64], %f21;
$L__BB0_34:
setp.gt.s32 %p53, %r2, 766;
or.pred %p54, %p1, %p53;
@%p54 bra $L__BB0_36;
ld.global.nc.f32 %f22, [%rd1+68];
st.shared.f32 [%r3+68], %f22;
$L__BB0_36:
setp.gt.s32 %p56, %r2, 765;
or.pred %p57, %p1, %p56;
@%p57 bra $L__BB0_38;
ld.global.nc.f32 %f23, [%rd1+72];
st.shared.f32 [%r3+72], %f23;
$L__BB0_38:
setp.gt.s32 %p59, %r2, 764;
or.pred %p60, %p1, %p59;
@%p60 bra $L__BB0_40;
ld.global.nc.f32 %f24, [%rd1+76];
st.shared.f32 [%r3+76], %f24;
$L__BB0_40:
setp.gt.s32 %p62, %r2, 763;
or.pred %p63, %p1, %p62;
@%p63 bra $L__BB0_42;
ld.global.nc.f32 %f25, [%rd1+80];
st.shared.f32 [%r3+80], %f25;
$L__BB0_42:
setp.gt.s32 %p65, %r2, 762;
or.pred %p66, %p1, %p65;
@%p66 bra $L__BB0_44;
ld.global.nc.f32 %f26, [%rd1+84];
st.shared.f32 [%r3+84], %f26;
$L__BB0_44:
setp.gt.s32 %p68, %r2, 761;
or.pred %p69, %p1, %p68;
@%p69 bra $L__BB0_46;
ld.global.nc.f32 %f27, [%rd1+88];
st.shared.f32 [%r3+88], %f27;
$L__BB0_46:
setp.gt.s32 %p71, %r2, 760;
or.pred %p72, %p1, %p71;
@%p72 bra $L__BB0_48;
ld.global.nc.f32 %f28, [%rd1+92];
st.shared.f32 [%r3+92], %f28;
$L__BB0_48:
setp.gt.s32 %p74, %r2, 759;
or.pred %p75, %p1, %p74;
@%p75 bra $L__BB0_50;
ld.global.nc.f32 %f29, [%rd1+96];
st.shared.f32 [%r3+96], %f29;
$L__BB0_50:
setp.gt.s32 %p77, %r2, 758;
or.pred %p78, %p1, %p77;
@%p78 bra $L__BB0_52;
ld.global.nc.f32 %f30, [%rd1+100];
st.shared.f32 [%r3+100], %f30;
$L__BB0_52:
setp.gt.s32 %p80, %r2, 757;
or.pred %p81, %p1, %p80;
@%p81 bra $L__BB0_54;
ld.global.nc.f32 %f31, [%rd1+104];
st.shared.f32 [%r3+104], %f31;
$L__BB0_54:
setp.gt.s32 %p83, %r2, 756;
or.pred %p84, %p1, %p83;
@%p84 bra $L__BB0_56;
ld.global.nc.f32 %f32, [%rd1+108];
st.shared.f32 [%r3+108], %f32;
$L__BB0_56:
cvta.to.global.u64 %rd13, %rd7;
cvta.to.global.u64 %rd2, %rd9;
for.size.1 i in range(0, 2, 1):
shl.b32 %r(17+i*1), %r1, (1+i*2);
mov.u32 %r19, _ZZ24default_function_kernel0E18placeholder_shared;
add.s32 %r20, %r19, %r18;
add.s32 %r90, %r20, 4096;
mul.wide.s32 %rd14, %r17, 4;
add.s64 %rd15, %rd13, %rd14;
add.s64 %rd19, %rd15, 4096;
cvta.to.global.u64 %rd4, %rd8;
mov.u32 %r91, 0;
$L__BB0_57:
for.size.2 i in range(0, 16, 1):
ld.global.nc.v2.u32 {%r(21+i*4), %r(22+i*4)}, [%rd19+(-4096+i*512)];
st.shared.v2.u32 [%r90+(-4096+i*512)], {%r(21+i*4), %r(22+i*4)};
add.s32 %r90, %r90, 8192;
add.s64 %rd19, %rd19, 8192;
add.s32 %r91, %r91, 16;
setp.ne.s32 %p85, %r91, 784;
@%p85 bra $L__BB0_57;
bar.sync 0;
mad.lo.s32 %r87, %r1, 3136, %r19;
add.s32 %r92, %r87, 200732;
mov.f32 %f80, 0f00000000;
mov.u32 %r93, 200704;
mov.f32 %f81, %f80;
$L__BB0_59:
add.s32 %r89, %r15, %r93;
for.size.5 i in range(0, 2, 1):
ld.shared.f32 %f(35+i*5), [%r92+(-200732+i*4)];
ld.shared.f32 %f(36+i*5), [%r89+(-200704+i*4)];
fma.rn.f32 %f(37+i*5), %f(36+i*5), %f(35+i*5), %f(80+i*-43);
ld.shared.f32 %f(38+i*5), [%r92+(-28+i*4)];
fma.rn.f32 %f(39+i*5), %f(36+i*5), %f(38+i*5), %f(81+i*-42);
for.size.5 i in range(0, 5, 1):
ld.shared.f32 %f(45+i*5), [%r92+(-200724+i*4)];
ld.shared.f32 %f(46+i*5), [%r89+(-200696+i*4)];
fma.rn.f32 %f(47+i*5), %f(46+i*5), %f(45+i*5), %f(42+i*5);
ld.shared.f32 %f(48+i*5), [%r92+(-20+i*4)];
fma.rn.f32 %f(49+i*5), %f(46+i*5), %f(48+i*5), %f(44+i*5);
ld.shared.f32 %f70, [%r92+-200704];
ld.shared.f32 %f71, [%r89+-200676];
fma.rn.f32 %f80, %f71, %f70, %f67;
ld.shared.f32 %f72, [%r92];
fma.rn.f32 %f81, %f71, %f72, %f69;
for.size.1 i in range(0, 2, 1):
add.s32 %r(92+i*1), %r(92+i*1), 32;
setp.ne.s32 %p86, %r93, 203840;
@%p86 bra $L__BB0_59;
mul.wide.s32 %rd16, %r1, 4;
add.s64 %rd17, %rd2, %rd16;
ld.global.nc.f32 %f73, [%rd17];
add.f32 %f74, %f80, %f73;
mov.f32 %f75, 0f00000000;
max.f32 %f76, %f74, %f75;
add.s64 %rd18, %rd4, %rd16;
st.global.f32 [%rd18], %f76;
ld.global.nc.f32 %f77, [%rd17+256];
add.f32 %f78, %f81, %f77;
max.f32 %f79, %f78, %f75;
st.global.f32 [%rd18+256], %f79;
ret;
}
| extern "C" __global__ void default_function_kernel0(float* __restrict__ placeholder, float* __restrict__ placeholder1, float* __restrict__ T_relu, float* __restrict__ placeholder2) {
float T_dense[2];
__shared__ float placeholder_d_shared[784];
__shared__ float placeholder_shared[100352];
T_dense[0] = 0.000000e+00f;
T_dense[1] = 0.000000e+00f;
for (int ax0_ax1_fused_inner_s = 0; ax0_ax1_fused_inner_s < 28; ++ax0_ax1_fused_inner_s) {
if ((int)threadIdx.x * 28 + ax0_ax1_fused_inner_s < 784) {
if ((int)threadIdx.x < 28) {
int index0 = (int)threadIdx.x * 28 + ax0_ax1_fused_inner_s;
placeholder_d_shared[index0] = placeholder[index0];
}
}
}
for (int ax0_ax1_fused_outer_outer = 0; ax0_ax1_fused_outer_outer < 784; ++ax0_ax1_fused_outer_outer) {
((float2*)(placeholder_shared + ax0_ax1_fused_outer_outer * 128 + (int)threadIdx.x * 2))[0] = ((float2*)(placeholder1 + ax0_ax1_fused_outer_outer * 128 + (int)threadIdx.x * 2))[0];
}
__syncthreads();
for (int k_outer_inner = 0; k_outer_inner < 98; ++k_outer_inner) {
for (int k_inner = 0; k_inner < 8; ++k_inner) {
int index1 = k_outer_inner * 8 + k_inner;
int index2 = (int)threadIdx.x * 784 + k_outer_inner * 8 + k_inner;
T_dense[0] = T_dense[0] + placeholder_d_shared[index1] * placeholder_shared[index2];
T_dense[1] = T_dense[1] + (placeholder_d_shared[index1] * placeholder_shared[index2 + 50176]);
}
}
int index3 = (int)threadIdx.x;
T_relu[index3] = max(T_dense[0] + placeholder2[index3], 0.000000e+00f);
T_relu[index3 + 64] = max(T_dense[1] + placeholder2[index3 + 64], 0.000000e+00f);
} |
[407f97a0b39db63cc31f3e635caadfd0,1,784,128,784,128,1,128] | tmppi9d78sx | default_function_kernel0 | .visible .entry default_function_kernel0(
.param .u64 default_function_kernel0_param_0,
.param .u64 default_function_kernel0_param_1,
.param .u64 default_function_kernel0_param_2,
.param .u64 default_function_kernel0_param_3
)
{
.reg .pred %p<3>;
.reg .f32 %f<643>;
.reg .b32 %r<1060>;
.reg .b64 %rd<18>;
// demoted variable
.shared .align 16 .b8 _ZZ24default_function_kernel0E20placeholder_d_shared[32];
// demoted variable
.shared .align 16 .b8 _ZZ24default_function_kernel0E18placeholder_shared[2048];
ld.param.u64 %rd1, [default_function_kernel0_param_0];
ld.param.u64 %rd2, [default_function_kernel0_param_1];
mov.u32 %r1058, 0;
mov.f32 %f579, 0f00000000;
cvta.to.global.u64 %rd5, %rd1;
mov.u32 %r24, %ctaid.x;
cvta.to.global.u64 %rd8, %rd2;
for.size.1 i in range(0, 63, 1):
mov.f32 %f(580+i*1), %f579;
$L__BB0_1:
mov.u32 %r1059, 0;
bar.sync 0;
shl.b32 %r7, %r1058, 3;
mul.wide.s32 %rd6, %r7, 4;
add.s64 %rd7, %rd5, %rd6;
for.size.2 i in range(0, 2, 1):
ld.global.nc.v4.u32 {%r(8+i*8), %r(9+i*8), %r(10+i*8), %r(11+i*8)}, [%rd7+(0+i*16)];
st.shared.v4.u32 [_ZZ24default_function_kernel0E20placeholder_d_shared+(0+i*16)], {%r(8+i*8), %r(9+i*8), %r(10+i*8), %r(11+i*8)};
mad.lo.s32 %r25, %r24, 50176, %r7;
mul.wide.s32 %rd9, %r25, 4;
add.s64 %rd10, %rd8, %rd9;
for.size.4 i in range(0, 64, 1):
for.size.1 j in range(0, 2, 1):
ld.global.nc.v4.u32 {%r(26+i*16+j*8), %r(27+i*16+j*8), %r(28+i*16+j*8), %r(29+i*16+j*8)}, [%rd10+(0+i*3136+j*16)];
for.size.1 j in range(0, 2, 1):
st.shared.v4.u32 [_ZZ24default_function_kernel0E18placeholder_shared+(0+i*32+j*16)], {%r(26+i*16+j*8), %r(27+i*16+j*8), %r(28+i*16+j*8), %r(29+i*16+j*8)};
bar.sync 0;
$L__BB0_2:
shl.b32 %r1050, %r1059, 2;
mov.u32 %r1051, _ZZ24default_function_kernel0E18placeholder_shared;
add.s32 %r1052, %r1051, %r1050;
ld.shared.f32 %f257, [%r1052];
mov.u32 %r1053, _ZZ24default_function_kernel0E20placeholder_d_shared;
add.s32 %r1054, %r1053, %r1050;
ld.shared.f32 %f258, [%r1054];
for.size.2 i in range(0, 2, 1):
fma.rn.f32 %f(642+i*-32), %f258, %f(257+i*2), %f(642+i*-32);
ld.shared.f32 %f(259+i*1), [%r1052+(1024+i*-992)];
for.size.3 i in range(0, 30, 1):
for.size.2 j in range(0, 2, 1):
fma.rn.f32 %f(641+i*-1+j*-32), %f258, %f(260+i*2+j*1), %f(641+i*-1+j*-32);
ld.shared.f32 %f(261+i*2+j*1), [%r1052+(1056+i*32+j*-992)];
fma.rn.f32 %f611, %f258, %f320, %f611;
ld.shared.f32 %f321, [%r1052+2016];
fma.rn.f32 %f579, %f258, %f321, %f579;
add.s32 %r1059, %r1059, 1;
setp.ne.s32 %p1, %r1059, 8;
@%p1 bra $L__BB0_2;
add.s32 %r1058, %r1058, 1;
setp.lt.u32 %p2, %r1058, 98;
@%p2 bra $L__BB0_1;
ld.param.u64 %rd17, [default_function_kernel0_param_2];
ld.param.u64 %rd16, [default_function_kernel0_param_3];
shl.b32 %r1056, %r24, 6;
cvta.to.global.u64 %rd11, %rd16;
mul.wide.s32 %rd12, %r1056, 4;
add.s64 %rd13, %rd11, %rd12;
ld.global.nc.f32 %f322, [%rd13];
add.f32 %f323, %f642, %f322;
mov.f32 %f324, 0f00000000;
max.f32 %f325, %f323, %f324;
cvta.to.global.u64 %rd14, %rd17;
add.s64 %rd15, %rd14, %rd12;
for.size.5 i in range(0, 31, 1):
for.size.4 j in range(0, 2, 1):
st.global.f32 [%rd15+(0+i*4+j*128)], %f(325+i*6+j*3);
ld.global.nc.f32 %f(326+i*6+j*3), [%rd13+(128+i*4+j*-124)];
add.f32 %f(327+i*6+j*3), %f(610+i*-1+j*31), %f(326+i*6+j*3);
max.f32 %f(328+i*6+j*3), %f(327+i*6+j*3), %f324;
st.global.f32 [%rd15+124], %f511;
ld.global.nc.f32 %f512, [%rd13+252];
add.f32 %f513, %f579, %f512;
max.f32 %f514, %f513, %f324;
st.global.f32 [%rd15+252], %f514;
ret;
}
| extern "C" __global__ void default_function_kernel0(float* __restrict__ placeholder, float* __restrict__ placeholder1, float* __restrict__ T_relu, float* __restrict__ placeholder2) {
float T_dense[64];
__shared__ float placeholder_d_shared[8];
__shared__ float placeholder_shared[512];
for (int j_inner_init = 0; j_inner_init < 32; ++j_inner_init) {
int index0 = j_inner_init;
T_dense[index0] = 0.000000e+00f;
T_dense[index0 + 32] = 0.000000e+00f;
}
for (int k_outer_outer = 0; k_outer_outer < 98; ++k_outer_outer) {
__syncthreads();
((float4*)(placeholder_d_shared + 0))[0] = ((float4*)(placeholder + k_outer_outer * 8))[0];
((float4*)(placeholder_d_shared + 4))[0] = ((float4*)(placeholder + k_outer_outer * 8 + 4))[0];
for (int ax0_ax1_fused_outer_outer = 0; ax0_ax1_fused_outer_outer < 64; ++ax0_ax1_fused_outer_outer) {
ulonglong4*(placeholder_shared + ax0_ax1_fused_outer_outer * 8)[0] = ulonglong4*(placeholder1 + (int)blockIdx.x * 50176 + ax0_ax1_fused_outer_outer * 784 + k_outer_outer * 8)[0];
}
__syncthreads();
for (int k_outer_inner = 0; k_outer_inner < 8; ++k_outer_inner) {
for (int j_inner = 0; j_inner < 32; ++j_inner) {
int index1 = j_inner;
int index2 = k_outer_inner;
int index3 = j_inner * 8 + k_outer_inner;
T_dense[index1] = T_dense[index1] + placeholder_d_shared[index2] * placeholder_shared[index3];
T_dense[index1 + 32] = T_dense[index1 + 32] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 256]);
}
}
}
for (int ax1_inner = 0; ax1_inner < 32; ++ax1_inner) {
int index4 = (int)blockIdx.x * 64 + ax1_inner;
int index5 = ax1_inner;
T_relu[index4] = max(T_dense[index5] + placeholder2[index4], 0.000000e+00f);
T_relu[index4 + 32] = max(T_dense[index5 + 32] + placeholder2[index4 + 32], 0.000000e+00f);
}
} |
[407f97a0b39db63cc31f3e635caadfd0,1,784,128,784,128,1,128] | tmpizw4ngdm | default_function_kernel0 | .visible .entry default_function_kernel0(
.param .u64 default_function_kernel0_param_0,
.param .u64 default_function_kernel0_param_1,
.param .u64 default_function_kernel0_param_2,
.param .u64 default_function_kernel0_param_3
)
{
.reg .pred %p<3>;
.reg .f32 %f<292>;
.reg .b32 %r<18>;
.reg .b64 %rd<23>;
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E20placeholder_d_shared[4];
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E18placeholder_shared[256];
ld.param.u64 %rd9, [default_function_kernel0_param_0];
ld.param.u64 %rd10, [default_function_kernel0_param_1];
ld.param.u64 %rd11, [default_function_kernel0_param_2];
ld.param.u64 %rd8, [default_function_kernel0_param_3];
cvta.to.global.u64 %rd12, %rd10;
cvta.to.global.u64 %rd1, %rd11;
mov.u32 %r1, %tid.x;
shl.b32 %r4, %r1, 2;
mov.u32 %r10, _ZZ24default_function_kernel0E20placeholder_d_shared;
add.s32 %r2, %r10, %r4;
mov.u32 %r11, _ZZ24default_function_kernel0E18placeholder_shared;
add.s32 %r3, %r11, %r4;
shl.b32 %r12, %r1, 4;
add.s32 %r5, %r11, %r12;
mul.lo.s32 %r13, %r1, 784;
cvta.to.global.u64 %rd13, %rd9;
mul.wide.s32 %rd14, %r13, 4;
add.s64 %rd22, %rd13, %rd14;
mov.u32 %r6, %ctaid.x;
mad.lo.s32 %r14, %r6, 50176, %r13;
mul.wide.s32 %rd15, %r14, 4;
add.s64 %rd16, %rd12, %rd15;
add.s64 %rd21, %rd16, 100352;
mov.u32 %r17, 0;
mov.f32 %f260, 0f00000000;
for.size.1 i in range(0, 31, 1):
mov.f32 %f(261+i*1), %f260;
$L__BB0_1:
bar.sync 0;
setp.gt.s32 %p1, %r1, 0;
@%p1 bra $L__BB0_3;
ld.global.nc.f32 %f97, [%rd22];
st.shared.f32 [%r2], %f97;
$L__BB0_3:
for.size.2 i in range(0, 32, 1):
ld.global.nc.f32 %f(98+i*1), [%rd21+(-100352+i*6272)];
st.shared.f32 [%r3+(0+i*8)], %f(98+i*1);
bar.sync 0;
ld.shared.f32 %f130, [%r5];
ld.shared.f32 %f131, [_ZZ24default_function_kernel0E20placeholder_d_shared];
for.size.2 i in range(0, 2, 1):
fma.rn.f32 %f(291+i*-1), %f131, %f(130+i*2), %f(291+i*-1);
ld.shared.f32 %f(132+i*1), [%r5+(32+i*32)];
for.size.2 i in range(0, 5, 1):
fma.rn.f32 %f(289+i*-1), %f131, %f(133+i*1), %f(289+i*-1);
ld.shared.f32 %f(134+i*1), [%r5+(96+i*32)];
for.size.3 i in range(0, 3, 1):
for.size.2 j in range(0, 8, 1):
fma.rn.f32 %f(284+i*-8+j*-1), %f131, %f(138+i*8+j*1), %f(284+i*-8+j*-1);
ld.shared.f32 %f(139+i*8+j*1), [%r5+(4+i*4+j*32)];
fma.rn.f32 %f260, %f131, %f162, %f260;
for.size.1 i in range(0, 2, 1):
add.s64 %rd(22+i*-1), %rd(22+i*-1), 4;
add.s32 %r17, %r17, 1;
setp.ne.s32 %p2, %r17, 784;
@%p2 bra $L__BB0_1;
cvta.to.global.u64 %rd17, %rd8;
shl.b32 %r15, %r6, 6;
add.s32 %r16, %r15, %r4;
mul.wide.s32 %rd18, %r16, 4;
add.s64 %rd19, %rd17, %rd18;
ld.global.nc.f32 %f163, [%rd19];
add.f32 %f164, %f291, %f163;
mov.f32 %f165, 0f00000000;
max.f32 %f166, %f164, %f165;
add.s64 %rd20, %rd1, %rd18;
for.size.4 i in range(0, 7, 1):
st.global.f32 [%rd20+(0+i*32)], %f(166+i*3);
ld.global.nc.f32 %f(167+i*3), [%rd19+(32+i*32)];
add.f32 %f(168+i*3), %f(290+i*-1), %f(167+i*3);
max.f32 %f(169+i*3), %f(168+i*3), %f165;
for.size.10 i in range(0, 3, 1):
for.size.4 j in range(0, 2, 1):
st.global.f32 [%rd20+(224+i*4+j*-220)], %f(187+i*24+j*3);
ld.global.nc.f32 %f(188+i*24+j*3), [%rd19+(4+i*4+j*32)];
add.f32 %f(189+i*24+j*3), %f(283+i*-8+j*-1), %f(188+i*24+j*3);
max.f32 %f(190+i*24+j*3), %f(189+i*24+j*3), %f165;
for.size.4 j in range(0, 6, 1):
st.global.f32 [%rd20+(36+i*4+j*32)], %f(193+i*24+j*3);
ld.global.nc.f32 %f(194+i*24+j*3), [%rd19+(68+i*4+j*32)];
add.f32 %f(195+i*24+j*3), %f(281+i*-8+j*-1), %f(194+i*24+j*3);
max.f32 %f(196+i*24+j*3), %f(195+i*24+j*3), %f165;
st.global.f32 [%rd20+236], %f259;
ret;
}
| extern "C" __global__ void default_function_kernel0(float* __restrict__ placeholder, float* __restrict__ placeholder1, float* __restrict__ T_relu, float* __restrict__ placeholder2) {
float T_dense[32];
__shared__ float placeholder_d_shared[1];
__shared__ float placeholder_shared[64];
T_dense[0] = 0.000000e+00f;
T_dense[4] = 0.000000e+00f;
T_dense[8] = 0.000000e+00f;
T_dense[12] = 0.000000e+00f;
T_dense[16] = 0.000000e+00f;
T_dense[20] = 0.000000e+00f;
T_dense[24] = 0.000000e+00f;
T_dense[28] = 0.000000e+00f;
T_dense[1] = 0.000000e+00f;
T_dense[5] = 0.000000e+00f;
T_dense[9] = 0.000000e+00f;
T_dense[13] = 0.000000e+00f;
T_dense[17] = 0.000000e+00f;
T_dense[21] = 0.000000e+00f;
T_dense[25] = 0.000000e+00f;
T_dense[29] = 0.000000e+00f;
T_dense[2] = 0.000000e+00f;
T_dense[6] = 0.000000e+00f;
T_dense[10] = 0.000000e+00f;
T_dense[14] = 0.000000e+00f;
T_dense[18] = 0.000000e+00f;
T_dense[22] = 0.000000e+00f;
T_dense[26] = 0.000000e+00f;
T_dense[30] = 0.000000e+00f;
T_dense[3] = 0.000000e+00f;
T_dense[7] = 0.000000e+00f;
T_dense[11] = 0.000000e+00f;
T_dense[15] = 0.000000e+00f;
T_dense[19] = 0.000000e+00f;
T_dense[23] = 0.000000e+00f;
T_dense[27] = 0.000000e+00f;
T_dense[31] = 0.000000e+00f;
for (int k_outer_outer = 0; k_outer_outer < 784; ++k_outer_outer) {
__syncthreads();
if ((int)threadIdx.x < 1) {
int index0 = (int)threadIdx.x;
int index1 = (int)threadIdx.x * 784 + k_outer_outer;
placeholder_d_shared[index0] = placeholder[index1];
}
int index0 = (int)threadIdx.x;
int index2 = (int)blockIdx.x * 50176 + (int)threadIdx.x * 784 + k_outer_outer;
placeholder_shared[index0] = placeholder1[index2];
placeholder_shared[index0 + 2] = placeholder1[index2 + 1568];
placeholder_shared[index0 + 4] = placeholder1[index2 + 3136];
placeholder_shared[index0 + 6] = placeholder1[index2 + 4704];
placeholder_shared[index0 + 8] = placeholder1[index2 + 6272];
placeholder_shared[index0 + 10] = placeholder1[index2 + 7840];
placeholder_shared[index0 + 12] = placeholder1[index2 + 9408];
placeholder_shared[index0 + 14] = placeholder1[index2 + 10976];
placeholder_shared[index0 + 16] = placeholder1[index2 + 12544];
placeholder_shared[index0 + 18] = placeholder1[index2 + 14112];
placeholder_shared[index0 + 20] = placeholder1[index2 + 15680];
placeholder_shared[index0 + 22] = placeholder1[index2 + 17248];
placeholder_shared[index0 + 24] = placeholder1[index2 + 18816];
placeholder_shared[index0 + 26] = placeholder1[index2 + 20384];
placeholder_shared[index0 + 28] = placeholder1[index2 + 21952];
placeholder_shared[index0 + 30] = placeholder1[index2 + 23520];
placeholder_shared[index0 + 32] = placeholder1[index2 + 25088];
placeholder_shared[index0 + 34] = placeholder1[index2 + 26656];
placeholder_shared[index0 + 36] = placeholder1[index2 + 28224];
placeholder_shared[index0 + 38] = placeholder1[index2 + 29792];
placeholder_shared[index0 + 40] = placeholder1[index2 + 31360];
placeholder_shared[index0 + 42] = placeholder1[index2 + 32928];
placeholder_shared[index0 + 44] = placeholder1[index2 + 34496];
placeholder_shared[index0 + 46] = placeholder1[index2 + 36064];
placeholder_shared[index0 + 48] = placeholder1[index2 + 37632];
placeholder_shared[index0 + 50] = placeholder1[index2 + 39200];
placeholder_shared[index0 + 52] = placeholder1[index2 + 40768];
placeholder_shared[index0 + 54] = placeholder1[index2 + 42336];
placeholder_shared[index0 + 56] = placeholder1[index2 + 43904];
placeholder_shared[index0 + 58] = placeholder1[index2 + 45472];
placeholder_shared[index0 + 60] = placeholder1[index2 + 47040];
placeholder_shared[index0 + 62] = placeholder1[index2 + 48608];
__syncthreads();
T_dense[0] = T_dense[0] + placeholder_d_shared[0] * placeholder_shared[index0 * 4];
int index3 = (int)threadIdx.x * 4 + 8;
T_dense[4] = T_dense[4] + placeholder_d_shared[0] * placeholder_shared[index3];
int index4 = (int)threadIdx.x * 4 + 16;
T_dense[8] = T_dense[8] + placeholder_d_shared[0] * placeholder_shared[index4];
int index5 = (int)threadIdx.x * 4 + 24;
T_dense[12] = T_dense[12] + placeholder_d_shared[0] * placeholder_shared[index5];
int index6 = (int)threadIdx.x * 4 + 32;
T_dense[16] = T_dense[16] + placeholder_d_shared[0] * placeholder_shared[index6];
int index7 = (int)threadIdx.x * 4 + 40;
T_dense[20] = T_dense[20] + placeholder_d_shared[0] * placeholder_shared[index7];
int index8 = (int)threadIdx.x * 4 + 48;
T_dense[24] = T_dense[24] + placeholder_d_shared[0] * placeholder_shared[index8];
int index9 = (int)threadIdx.x * 4 + 56;
T_dense[28] = T_dense[28] + placeholder_d_shared[0] * placeholder_shared[index9];
int index10 = (int)threadIdx.x * 4 + 1;
T_dense[1] = T_dense[1] + placeholder_d_shared[0] * placeholder_shared[index10];
int index11 = (int)threadIdx.x * 4 + 9;
T_dense[5] = T_dense[5] + placeholder_d_shared[0] * placeholder_shared[index11];
int index12 = (int)threadIdx.x * 4 + 17;
T_dense[9] = T_dense[9] + placeholder_d_shared[0] * placeholder_shared[index12];
int index13 = (int)threadIdx.x * 4 + 25;
T_dense[13] = T_dense[13] + placeholder_d_shared[0] * placeholder_shared[index13];
int index14 = (int)threadIdx.x * 4 + 33;
T_dense[17] = T_dense[17] + placeholder_d_shared[0] * placeholder_shared[index14];
int index15 = (int)threadIdx.x * 4 + 41;
T_dense[21] = T_dense[21] + placeholder_d_shared[0] * placeholder_shared[index15];
int index16 = (int)threadIdx.x * 4 + 49;
T_dense[25] = T_dense[25] + placeholder_d_shared[0] * placeholder_shared[index16];
int index17 = (int)threadIdx.x * 4 + 57;
T_dense[29] = T_dense[29] + placeholder_d_shared[0] * placeholder_shared[index17];
int index18 = (int)threadIdx.x * 4 + 2;
T_dense[2] = T_dense[2] + placeholder_d_shared[0] * placeholder_shared[index18];
int index19 = (int)threadIdx.x * 4 + 10;
T_dense[6] = T_dense[6] + placeholder_d_shared[0] * placeholder_shared[index19];
int index20 = (int)threadIdx.x * 4 + 18;
T_dense[10] = T_dense[10] + placeholder_d_shared[0] * placeholder_shared[index20];
int index21 = (int)threadIdx.x * 4 + 26;
T_dense[14] = T_dense[14] + placeholder_d_shared[0] * placeholder_shared[index21];
int index22 = (int)threadIdx.x * 4 + 34;
T_dense[18] = T_dense[18] + placeholder_d_shared[0] * placeholder_shared[index22];
int index23 = (int)threadIdx.x * 4 + 42;
T_dense[22] = T_dense[22] + placeholder_d_shared[0] * placeholder_shared[index23];
int index24 = (int)threadIdx.x * 4 + 50;
T_dense[26] = T_dense[26] + placeholder_d_shared[0] * placeholder_shared[index24];
int index25 = (int)threadIdx.x * 4 + 58;
T_dense[30] = T_dense[30] + placeholder_d_shared[0] * placeholder_shared[index25];
int index26 = (int)threadIdx.x * 4 + 3;
T_dense[3] = T_dense[3] + placeholder_d_shared[0] * placeholder_shared[index26];
int index27 = (int)threadIdx.x * 4 + 11;
T_dense[7] = T_dense[7] + placeholder_d_shared[0] * placeholder_shared[index27];
int index28 = (int)threadIdx.x * 4 + 19;
T_dense[11] = T_dense[11] + placeholder_d_shared[0] * placeholder_shared[index28];
int index29 = (int)threadIdx.x * 4 + 27;
T_dense[15] = T_dense[15] + placeholder_d_shared[0] * placeholder_shared[index29];
int index30 = (int)threadIdx.x * 4 + 35;
T_dense[19] = T_dense[19] + placeholder_d_shared[0] * placeholder_shared[index30];
int index31 = (int)threadIdx.x * 4 + 43;
T_dense[23] = T_dense[23] + placeholder_d_shared[0] * placeholder_shared[index31];
int index32 = (int)threadIdx.x * 4 + 51;
T_dense[27] = T_dense[27] + placeholder_d_shared[0] * placeholder_shared[index32];
int index33 = (int)threadIdx.x * 4 + 59;
T_dense[31] = T_dense[31] + placeholder_d_shared[0] * placeholder_shared[index33];
}
for (int ax1_inner = 0; ax1_inner < 4; ++ax1_inner) {
int index34 = (int)blockIdx.x * 64 + (int)threadIdx.x * 4 + ax1_inner;
int index35 = ax1_inner;
T_relu[index34] = max(T_dense[index35] + placeholder2[index34], 0.000000e+00f);
T_relu[index34 + 8] = max(T_dense[index35 + 4] + placeholder2[index34 + 8], 0.000000e+00f);
T_relu[index34 + 16] = max(T_dense[index35 + 8] + placeholder2[index34 + 16], 0.000000e+00f);
T_relu[index34 + 24] = max(T_dense[index35 + 12] + placeholder2[index34 + 24], 0.000000e+00f);
T_relu[index34 + 32] = max(T_dense[index35 + 16] + placeholder2[index34 + 32], 0.000000e+00f);
T_relu[index34 + 40] = max(T_dense[index35 + 20] + placeholder2[index34 + 40], 0.000000e+00f);
T_relu[index34 + 48] = max(T_dense[index35 + 24] + placeholder2[index34 + 48], 0.000000e+00f);
T_relu[index34 + 56] = max(T_dense[index35 + 28] + placeholder2[index34 + 56], 0.000000e+00f);
}
} |
[407f97a0b39db63cc31f3e635caadfd0,1,784,128,784,128,1,128] | tmp7wub3ptp | default_function_kernel0 | .visible .entry default_function_kernel0(
.param .u64 default_function_kernel0_param_0,
.param .u64 default_function_kernel0_param_1,
.param .u64 default_function_kernel0_param_2,
.param .u64 default_function_kernel0_param_3
)
{
.reg .pred %p<5>;
.reg .f32 %f<229>;
.reg .b32 %r<108>;
.reg .b64 %rd<30>;
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E20placeholder_d_shared[784];
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E18placeholder_shared[100352];
ld.param.u64 %rd6, [default_function_kernel0_param_0];
ld.param.u64 %rd7, [default_function_kernel0_param_1];
ld.param.u64 %rd8, [default_function_kernel0_param_2];
ld.param.u64 %rd9, [default_function_kernel0_param_3];
cvta.to.global.u64 %rd1, %rd7;
cvta.to.global.u64 %rd2, %rd6;
cvta.to.global.u64 %rd3, %rd8;
cvta.to.global.u64 %rd4, %rd9;
mov.u32 %r1, %tid.x;
shl.b32 %r18, %r1, 2;
mov.u32 %r19, _ZZ24default_function_kernel0E20placeholder_d_shared;
add.s32 %r2, %r19, %r18;
mul.lo.s32 %r3, %r1, 3136;
add.s32 %r4, %r1, 56;
mov.u32 %r20, _ZZ24default_function_kernel0E18placeholder_shared;
add.s32 %r5, %r20, %r18;
mov.u32 %r103, 0;
mov.f32 %f197, 0f00000000;
for.size.1 i in range(0, 15, 1):
mov.f32 %f(198+i*1), %f197;
$L__BB0_1:
bar.sync 0;
mul.lo.s32 %r7, %r103, 196;
add.s32 %r21, %r7, %r1;
mul.wide.s32 %rd10, %r21, 4;
add.s64 %rd5, %rd2, %rd10;
for.size.2 i in range(0, 24, 1):
ld.global.nc.f32 %f(65+i*1), [%rd5+(0+i*32)];
st.shared.f32 [%r2+(0+i*32)], %f(65+i*1);
setp.gt.s32 %p1, %r1, 3;
@%p1 bra $L__BB0_3;
ld.global.nc.f32 %f89, [%rd5+768];
st.shared.f32 [%r2+768], %f89;
$L__BB0_3:
mov.u32 %r106, 128;
mov.u32 %r104, %r1;
mov.u32 %r105, %r4;
$L__BB0_4:
mul.hi.s32 %r23, %r104, 1402438301;
shr.u32 %r24, %r23, 31;
shr.s32 %r25, %r23, 6;
add.s32 %r26, %r25, %r24;
mul.lo.s32 %r27, %r26, 196;
sub.s32 %r28, %r104, %r27;
add.s32 %r29, %r28, %r7;
mad.lo.s32 %r30, %r26, 784, %r29;
mul.wide.s32 %rd11, %r30, 4;
add.s64 %rd12, %rd1, %rd11;
ld.global.nc.f32 %f90, [%rd12];
add.s32 %r31, %r5, %r106;
for.size.13 i in range(0, 4, 1):
st.shared.f32 [%r31+(-128+i*32)], %f(90+i*1);
add.s32 %r(32+i*9), %r104, (8+i*8);
mul.hi.s32 %r(33+i*9), %r(32+i*9), 1402438301;
shr.u32 %r(34+i*9), %r(33+i*9), 31;
shr.s32 %r(35+i*9), %r(33+i*9), 6;
add.s32 %r(36+i*9), %r(35+i*9), %r(34+i*9);
mul.lo.s32 %r(37+i*9), %r(36+i*9), 196;
sub.s32 %r(38+i*9), %r(32+i*9), %r(37+i*9);
add.s32 %r(39+i*9), %r(38+i*9), %r7;
mad.lo.s32 %r(40+i*9), %r(36+i*9), 784, %r(39+i*9);
mul.wide.s32 %rd(13+i*2), %r(40+i*9), 4;
add.s64 %rd(14+i*2), %rd1, %rd(13+i*2);
ld.global.nc.f32 %f(91+i*1), [%rd(14+i*2)+0];
for.size.13 i in range(0, 2, 1):
st.shared.f32 [%r31+(0+i*32)], %f(94+i*1);
add.s32 %r(68+i*9), %r105, (-16+i*8);
mul.hi.s32 %r(69+i*9), %r(68+i*9), 1402438301;
shr.u32 %r(70+i*9), %r(69+i*9), 31;
shr.s32 %r(71+i*9), %r(69+i*9), 6;
add.s32 %r(72+i*9), %r(71+i*9), %r(70+i*9);
mul.lo.s32 %r(73+i*9), %r(72+i*9), 196;
sub.s32 %r(74+i*9), %r(68+i*9), %r(73+i*9);
add.s32 %r(75+i*9), %r(74+i*9), %r7;
mad.lo.s32 %r(76+i*9), %r(72+i*9), 784, %r(75+i*9);
mul.wide.s32 %rd(21+i*2), %r(76+i*9), 4;
add.s64 %rd(22+i*2), %rd1, %rd(21+i*2);
ld.global.nc.f32 %f(95+i*1), [%rd(22+i*2)+0];
st.shared.f32 [%r31+64], %f96;
mul.hi.s32 %r86, %r105, 1402438301;
shr.u32 %r87, %r86, 31;
shr.s32 %r88, %r86, 6;
add.s32 %r89, %r88, %r87;
mul.lo.s32 %r90, %r89, 196;
sub.s32 %r91, %r105, %r90;
add.s32 %r92, %r91, %r7;
mad.lo.s32 %r93, %r89, 784, %r92;
mul.wide.s32 %rd25, %r93, 4;
add.s64 %rd26, %rd1, %rd25;
ld.global.nc.f32 %f97, [%rd26];
st.shared.f32 [%r31+96], %f97;
add.s32 %r106, %r106, 256;
setp.ne.s32 %p2, %r106, 100480;
for.size.1 i in range(0, 2, 1):
add.s32 %r(12+i*92), %r105, (64+i*-56);
mov.u32 %r105, %r12;
@%p2 bra $L__BB0_4;
bar.sync 0;
mov.u32 %r107, 0;
$L__BB0_6:
add.s32 %r95, %r107, %r3;
for.size.3 i in range(0, 2, 1):
shl.b32 %r(96+i*3), %r(95+i*12), 2;
add.s32 %r(98+i*3), %r(20+i*-1), %r(96+i*3);
ld.shared.f32 %f(98+i*1), [%r(98+i*3)+0];
for.size.2 i in range(0, 2, 1):
fma.rn.f32 %f(100+i*2), %f99, %f(98+i*3), %f(212+i*-1);
ld.shared.f32 %f(101+i*2), [%r98+(784+i*784)];
for.size.2 i in range(0, 13, 1):
fma.rn.f32 %f(104+i*2), %f99, %f(103+i*2), %f(210+i*-1);
ld.shared.f32 %f(105+i*2), [%r98+(2352+i*784)];
fma.rn.f32 %f130, %f99, %f129, %f197;
for.size.1 i in range(0, 2, 1):
ld.shared.f32 %f(131+i*1), [%r(98+i*3)+4];
for.size.2 i in range(0, 2, 1):
fma.rn.f32 %f(212+i*-1), %f132, %f(131+i*2), %f(100+i*2);
ld.shared.f32 %f(133+i*1), [%r98+(788+i*784)];
for.size.2 i in range(0, 13, 1):
fma.rn.f32 %f(210+i*-1), %f132, %f(134+i*1), %f(104+i*2);
ld.shared.f32 %f(135+i*1), [%r98+(2356+i*784)];
fma.rn.f32 %f197, %f132, %f147, %f130;
add.s32 %r107, %r107, 2;
setp.ne.s32 %p3, %r107, 196;
@%p3 bra $L__BB0_6;
add.s32 %r103, %r103, 1;
setp.lt.u32 %p4, %r103, 4;
@%p4 bra $L__BB0_1;
shl.b32 %r102, %r1, 4;
mul.wide.s32 %rd27, %r102, 4;
add.s64 %rd28, %rd4, %rd27;
ld.global.nc.f32 %f148, [%rd28];
add.f32 %f149, %f212, %f148;
mov.f32 %f150, 0f00000000;
max.f32 %f151, %f149, %f150;
add.s64 %rd29, %rd3, %rd27;
for.size.4 i in range(0, 15, 1):
st.global.f32 [%rd29+(0+i*4)], %f(151+i*3);
ld.global.nc.f32 %f(152+i*3), [%rd28+(4+i*4)];
add.f32 %f(153+i*3), %f(211+i*-1), %f(152+i*3);
max.f32 %f(154+i*3), %f(153+i*3), %f150;
st.global.f32 [%rd29+60], %f196;
ret;
}
| extern "C" __global__ void default_function_kernel0(float* __restrict__ placeholder, float* __restrict__ placeholder1, float* __restrict__ T_relu, float* __restrict__ placeholder2) {
float T_dense[16];
__shared__ float placeholder_d_shared[196];
__shared__ float placeholder_shared[25088];
T_dense[0] = 0.000000e+00f;
T_dense[1] = 0.000000e+00f;
T_dense[2] = 0.000000e+00f;
T_dense[3] = 0.000000e+00f;
T_dense[4] = 0.000000e+00f;
T_dense[5] = 0.000000e+00f;
T_dense[6] = 0.000000e+00f;
T_dense[7] = 0.000000e+00f;
T_dense[8] = 0.000000e+00f;
T_dense[9] = 0.000000e+00f;
T_dense[10] = 0.000000e+00f;
T_dense[11] = 0.000000e+00f;
T_dense[12] = 0.000000e+00f;
T_dense[13] = 0.000000e+00f;
T_dense[14] = 0.000000e+00f;
T_dense[15] = 0.000000e+00f;
for (int k_outer_outer = 0; k_outer_outer < 4; ++k_outer_outer) {
__syncthreads();
int index0 = (int)threadIdx.x;
int index1 = k_outer_outer * 196 + (int)threadIdx.x;
placeholder_d_shared[index0] = placeholder[index1];
placeholder_d_shared[index0 + 8] = placeholder[index1 + 8];
placeholder_d_shared[index0 + 16] = placeholder[index1 + 16];
placeholder_d_shared[index0 + 24] = placeholder[index1 + 24];
placeholder_d_shared[index0 + 32] = placeholder[index1 + 32];
placeholder_d_shared[index0 + 40] = placeholder[index1 + 40];
placeholder_d_shared[index0 + 48] = placeholder[index1 + 48];
placeholder_d_shared[index0 + 56] = placeholder[index1 + 56];
placeholder_d_shared[index0 + 64] = placeholder[index1 + 64];
placeholder_d_shared[index0 + 72] = placeholder[index1 + 72];
placeholder_d_shared[index0 + 80] = placeholder[index1 + 80];
placeholder_d_shared[index0 + 88] = placeholder[index1 + 88];
placeholder_d_shared[index0 + 96] = placeholder[index1 + 96];
placeholder_d_shared[index0 + 104] = placeholder[index1 + 104];
placeholder_d_shared[index0 + 112] = placeholder[index1 + 112];
placeholder_d_shared[index0 + 120] = placeholder[index1 + 120];
placeholder_d_shared[index0 + 128] = placeholder[index1 + 128];
placeholder_d_shared[index0 + 136] = placeholder[index1 + 136];
placeholder_d_shared[index0 + 144] = placeholder[index1 + 144];
placeholder_d_shared[index0 + 152] = placeholder[index1 + 152];
placeholder_d_shared[index0 + 160] = placeholder[index1 + 160];
placeholder_d_shared[index0 + 168] = placeholder[index1 + 168];
placeholder_d_shared[index0 + 176] = placeholder[index1 + 176];
placeholder_d_shared[index0 + 184] = placeholder[index1 + 184];
if ((int)threadIdx.x < 4) {
placeholder_d_shared[index0 + 192] = placeholder[index1 + 192];
}
for (int ax0_ax1_fused_outer_outer = 0; ax0_ax1_fused_outer_outer < 3136; ++ax0_ax1_fused_outer_outer) {
int index2 = ax0_ax1_fused_outer_outer * 8 + (int)threadIdx.x;
int index3 = (ax0_ax1_fused_outer_outer * 8 + (int)threadIdx.x) / 196 * 784 + k_outer_outer * 196 + (ax0_ax1_fused_outer_outer * 8 + (int)threadIdx.x) % 196;
placeholder_shared[index2] = placeholder1[index3];
}
__syncthreads();
for (int k_outer_inner = 0; k_outer_inner < 196; ++k_outer_inner) {
int index4 = k_outer_inner;
int index5 = (int)threadIdx.x * 3136 + k_outer_inner;
T_dense[0] = T_dense[0] + placeholder_d_shared[index4] * placeholder_shared[index5];
T_dense[1] = T_dense[1] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 196]);
T_dense[2] = T_dense[2] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 392]);
T_dense[3] = T_dense[3] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 588]);
T_dense[4] = T_dense[4] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 784]);
T_dense[5] = T_dense[5] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 980]);
T_dense[6] = T_dense[6] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1176]);
T_dense[7] = T_dense[7] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1372]);
T_dense[8] = T_dense[8] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1568]);
T_dense[9] = T_dense[9] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1764]);
T_dense[10] = T_dense[10] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1960]);
T_dense[11] = T_dense[11] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 2156]);
T_dense[12] = T_dense[12] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 2352]);
T_dense[13] = T_dense[13] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 2548]);
T_dense[14] = T_dense[14] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 2744]);
T_dense[15] = T_dense[15] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 2940]);
}
}
for (int ax1_inner = 0; ax1_inner < 16; ++ax1_inner) {
int index6 = (int)threadIdx.x * 16 + ax1_inner;
int index7 = ax1_inner;
T_relu[index6] = max(T_dense[index7] + placeholder2[index6], 0.000000e+00f);
}
} |
[407f97a0b39db63cc31f3e635caadfd0,1,784,128,784,128,1,128] | tmptrn7w2g4 | default_function_kernel0 | .visible .entry default_function_kernel0(
.param .u64 default_function_kernel0_param_0,
.param .u64 default_function_kernel0_param_1,
.param .u64 default_function_kernel0_param_2,
.param .u64 default_function_kernel0_param_3
)
{
.reg .pred %p<5>;
.reg .f32 %f<68>;
.reg .b32 %r<20>;
.reg .b64 %rd<34>;
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E7T_dense[8];
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E8red_buf0[8];
ld.param.u64 %rd16, [default_function_kernel0_param_0];
ld.param.u64 %rd17, [default_function_kernel0_param_1];
ld.param.u64 %rd18, [default_function_kernel0_param_2];
ld.param.u64 %rd19, [default_function_kernel0_param_3];
cvta.to.global.u64 %rd1, %rd17;
cvta.to.global.u64 %rd2, %rd18;
cvta.to.global.u64 %rd3, %rd19;
mov.u32 %r1, %tid.x;
shl.b32 %r9, %r1, 2;
mov.u32 %r10, _ZZ24default_function_kernel0E8red_buf0;
add.s32 %r2, %r10, %r9;
cvta.to.global.u64 %rd20, %rd16;
mul.wide.s32 %rd21, %r1, 4;
add.s64 %rd22, %rd20, %rd21;
add.s64 %rd33, %rd22, 32;
mov.u32 %r3, %ctaid.x;
mad.lo.s32 %r11, %r3, 1568, %r1;
cvt.s64.s32 %rd5, %r11;
mul.wide.s32 %rd23, %r11, 4;
add.s64 %rd24, %rd1, %rd23;
add.s64 %rd30, %rd24, 32;
mov.f32 %f66, 0f00000000;
mov.u32 %r18, 0;
mov.u64 %rd31, %rd33;
$L__BB0_1:
for.size.1 i in range(0, 2, 1):
ld.global.nc.f32 %f(6+i*1), [%rd(30+i*1)+-32];
for.size.3 i in range(0, 2, 1):
fma.rn.f32 %f(8+i*3), %f(7+i*3), %f(6+i*3), %f(66+i*-58);
ld.global.nc.f32 %f(9+i*3), [%rd30+(-24+i*8)];
ld.global.nc.f32 %f(10+i*3), [%rd31+(-24+i*8)];
for.size.3 i in range(0, 5, 1):
fma.rn.f32 %f(14+i*3), %f(13+i*3), %f(12+i*3), %f(11+i*3);
ld.global.nc.f32 %f(15+i*3), [%rd30+(-8+i*8)];
ld.global.nc.f32 %f(16+i*3), [%rd31+(-8+i*8)];
fma.rn.f32 %f66, %f28, %f27, %f26;
for.size.1 i in range(0, 2, 1):
add.s64 %rd(31+i*-1), %rd(31+i*-1), 64;
add.s32 %r18, %r18, 8;
setp.ne.s32 %p1, %r18, 392;
@%p1 bra $L__BB0_1;
setp.gt.s32 %p2, %r1, 0;
bar.sync 0;
st.volatile.shared.f32 [%r2], %f66;
bar.sync 0;
@%p2 bra $L__BB0_4;
for.size.1 i in range(0, 2, 1):
ld.volatile.shared.f32 %f(29+i*1), [%r2+(0+i*4)];
add.f32 %f31, %f29, %f30;
st.volatile.shared.f32 [%r2], %f31;
$L__BB0_4:
bar.sync 0;
mov.u32 %r19, 0;
ld.volatile.shared.f32 %f33, [_ZZ24default_function_kernel0E8red_buf0];
st.shared.f32 [_ZZ24default_function_kernel0E7T_dense], %f33;
shl.b64 %rd25, %rd5, 2;
add.s64 %rd26, %rd1, %rd25;
add.s64 %rd32, %rd26, 3192;
mov.f32 %f67, 0f00000000;
$L__BB0_5:
for.size.3 i in range(0, 2, 1):
ld.global.nc.f32 %f(34+i*3), [%rd32+(-56+i*8)];
ld.global.nc.f32 %f(35+i*3), [%rd33+(-32+i*8)];
fma.rn.f32 %f(36+i*3), %f(35+i*3), %f(34+i*3), %f(67+i*-31);
for.size.3 i in range(0, 5, 1):
ld.global.nc.f32 %f(40+i*3), [%rd32+(-40+i*8)];
ld.global.nc.f32 %f(41+i*3), [%rd33+(-16+i*8)];
fma.rn.f32 %f(42+i*3), %f(41+i*3), %f(40+i*3), %f(39+i*3);
ld.global.nc.f32 %f55, [%rd32];
ld.global.nc.f32 %f56, [%rd33+24];
fma.rn.f32 %f67, %f56, %f55, %f54;
for.size.1 i in range(0, 2, 1):
add.s64 %rd(33+i*-1), %rd(33+i*-1), 64;
add.s32 %r19, %r19, 8;
setp.ne.s32 %p3, %r19, 392;
@%p3 bra $L__BB0_5;
bar.sync 0;
st.volatile.shared.f32 [%r2], %f67;
bar.sync 0;
@%p2 bra $L__BB0_8;
for.size.1 i in range(0, 2, 1):
ld.volatile.shared.f32 %f(57+i*1), [%r2+(0+i*4)];
add.f32 %f59, %f57, %f58;
st.volatile.shared.f32 [%r2], %f59;
$L__BB0_8:
bar.sync 0;
ld.volatile.shared.f32 %f60, [_ZZ24default_function_kernel0E8red_buf0];
mov.u32 %r13, _ZZ24default_function_kernel0E7T_dense;
st.shared.f32 [_ZZ24default_function_kernel0E7T_dense+4], %f60;
bar.sync 0;
shl.b32 %r14, %r3, 1;
add.s32 %r15, %r14, %r1;
mul.wide.s32 %rd27, %r15, 4;
add.s64 %rd28, %rd3, %rd27;
ld.global.nc.f32 %f61, [%rd28];
add.s32 %r17, %r13, %r9;
ld.shared.f32 %f62, [%r17];
add.f32 %f63, %f62, %f61;
mov.f32 %f64, 0f00000000;
max.f32 %f65, %f63, %f64;
add.s64 %rd29, %rd2, %rd27;
st.global.f32 [%rd29], %f65;
ret;
}
| extern "C" __global__ void default_function_kernel0(float* __restrict__ placeholder, float* __restrict__ placeholder1, float* __restrict__ T_relu, float* __restrict__ placeholder2) {
__shared__ float T_dense[2];
float normal_reduce_temp0[1];
__shared__ float red_buf0[2];
for (int j = 0; j < 2; ++j) {
normal_reduce_temp0[0] = 0.000000e+00f;
for (int k_outer = 0; k_outer < 392; ++k_outer) {
int index0 = k_outer * 2 + (int)threadIdx.x;
int index1 = (int)blockIdx.x * 1568 + j * 784 + k_outer * 2 + (int)threadIdx.x;
normal_reduce_temp0[0] = normal_reduce_temp0[0] + placeholder[index0] * placeholder1[index1];
}
__syncthreads();
int index2 = (int)threadIdx.x;
((volatile float*)red_buf0)[index2] = normal_reduce_temp0[0];
__syncthreads();
if ((int)threadIdx.x < 1) {
((volatile float*)red_buf0)[index2] = ((volatile float*)red_buf0)[index2] + ((volatile float*)red_buf0)[index2 + 1];
}
__syncthreads();
int index3 = j;
T_dense[index3] = ((volatile float*)red_buf0)[0];
}
__syncthreads();
int index2 = (int)threadIdx.x;
int index4 = (int)blockIdx.x * 2 + (int)threadIdx.x;
T_relu[index4] = max(T_dense[index2] + placeholder2[index4], 0.000000e+00f);
} |
[407f97a0b39db63cc31f3e635caadfd0,1,784,128,784,128,1,128] | tmpt0gfnz24 | default_function_kernel0 | .visible .entry default_function_kernel0(
.param .u64 default_function_kernel0_param_0,
.param .u64 default_function_kernel0_param_1,
.param .u64 default_function_kernel0_param_2,
.param .u64 default_function_kernel0_param_3
)
{
.reg .pred %p<5>;
.reg .f32 %f<72>;
.reg .b32 %r<103>;
.reg .b64 %rd<18>;
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E20placeholder_d_shared[1568];
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E18placeholder_shared[100352];
ld.param.u64 %rd6, [default_function_kernel0_param_0];
ld.param.u64 %rd7, [default_function_kernel0_param_1];
ld.param.u64 %rd4, [default_function_kernel0_param_2];
ld.param.u64 %rd5, [default_function_kernel0_param_3];
cvta.to.global.u64 %rd1, %rd7;
cvta.to.global.u64 %rd2, %rd6;
mov.u32 %r20, %tid.x;
shl.b32 %r1, %r20, 1;
mov.u32 %r21, %ctaid.x;
for.size.1 i in range(0, 2, 1):
mul.lo.s32 %r(2+i*1), %r(21+i*-1), (50176+i*-49392);
shl.b32 %r22, %r20, 3;
add.s32 %r4, %r22, 256;
shl.b32 %r23, %r20, 5;
mov.u32 %r24, _ZZ24default_function_kernel0E18placeholder_shared;
add.s32 %r25, %r24, %r23;
add.s32 %r5, %r25, 1040;
mov.u32 %r98, 0;
mov.f32 %f68, 0f00000000;
mov.f32 %f69, %f68;
$L__BB0_1:
setp.gt.s32 %p1, %r20, 3;
bar.sync 0;
mul.lo.s32 %r7, %r98, 392;
add.s32 %r28, %r7, %r1;
mul.wide.s32 %rd8, %r28, 4;
add.s64 %rd3, %rd2, %rd8;
ld.global.nc.f32 %f9, [%rd3];
mov.u32 %r30, _ZZ24default_function_kernel0E20placeholder_d_shared;
add.s32 %r31, %r30, %r22;
for.size.3 i in range(0, 5, 1):
for.size.2 j in range(0, 2, 1):
st.shared.f32 [%r31+(0+i*256+j*4)], %f(9+i*2+j*1);
ld.global.nc.f32 %f(10+i*2+j*1), [%rd3+(4+i*256+j*252)];
st.shared.f32 [%r31+1280], %f19;
ld.global.nc.f32 %f20, [%rd3+1284];
st.shared.f32 [%r31+1284], %f20;
@%p1 bra $L__BB0_3;
for.size.2 i in range(0, 2, 1):
ld.global.nc.f32 %f(21+i*1), [%rd3+(1536+i*4)];
st.shared.f32 [%r31+(1536+i*4)], %f(21+i*1);
$L__BB0_3:
add.s32 %r9, %r7, %r2;
mov.u32 %r101, 1024;
mov.u32 %r99, %r5;
mov.u32 %r100, %r4;
$L__BB0_4:
add.s32 %r37, %r100, -256;
for.size.14 i in range(0, 2, 1):
mul.hi.s32 %r(38+i*24), %r(37+i*63), 1402438301;
shr.u32 %r(39+i*24), %r(38+i*24), 31;
shr.s32 %r(40+i*24), %r(38+i*24), 7;
add.s32 %r(41+i*24), %r(40+i*24), %r(39+i*24);
mul.lo.s32 %r(42+i*24), %r(41+i*24), 392;
sub.s32 %r(43+i*24), %r(37+i*63), %r(42+i*24);
add.s32 %r(44+i*24), %r9, %r(43+i*24);
mad.lo.s32 %r(45+i*24), %r(41+i*24), 784, %r(44+i*24);
mul.wide.s32 %rd(9+i*2), %r(45+i*24), 4;
add.s64 %rd(10+i*2), %rd1, %rd(9+i*2);
ld.global.nc.v4.u32 {%r(46+i*24), %r(47+i*24), %r(48+i*24), %r(49+i*24)}, [%rd(10+i*2)+0];
ld.global.nc.v4.u32 {%r(54+i*24), %r(55+i*24), %r(56+i*24), %r(57+i*24)}, [%rd(10+i*2)+16];
st.shared.v4.u32 [%r99+(-1040+i*1024)], {%r(46+i*24), %r(47+i*24), %r(48+i*24), %r(49+i*24)};
st.shared.v4.u32 [%r99+(-1024+i*1024)], {%r(54+i*24), %r(55+i*24), %r(56+i*24), %r(57+i*24)};
for.size.1 i in range(0, 2, 1):
add.s32 %r(100+i*-1), %r(100+i*-1), (512+i*1536);
add.s32 %r101, %r101, 2048;
setp.ne.s32 %p2, %r101, 101376;
@%p2 bra $L__BB0_4;
bar.sync 0;
mov.u32 %r102, 0;
$L__BB0_6:
for.size.1 i in range(0, 2, 1):
shl.b32 %r(87+i*1), %r102, (1+i*2);
for.size.1 i in range(0, 2, 1):
add.s32 %r(90+i*1), %r(30+i*57), %r(88+i*-85);
shl.b32 %r92, %r91, 2;
add.s32 %r94, %r24, %r92;
for.size.1 i in range(0, 2, 1):
ld.shared.f32 %f(23+i*1), [%r(94+i*-4)+0];
for.size.10 i in range(0, 2, 1):
fma.rn.f32 %f(25+i*10), %f(24+i*10), %f(23+i*10), %f(69+i*-41);
ld.shared.f32 %f(26+i*10), [%r94+(4+i*8)];
ld.shared.f32 %f(27+i*10), [%r90+(4+i*8)];
fma.rn.f32 %f(28+i*10), %f(27+i*10), %f(26+i*10), %f(25+i*10);
ld.shared.f32 %f(29+i*10), [%r94+(1568+i*8)];
fma.rn.f32 %f(30+i*10), %f(24+i*10), %f(29+i*10), %f(68+i*-36);
ld.shared.f32 %f(31+i*10), [%r94+(1572+i*8)];
fma.rn.f32 %f(32+i*10), %f(27+i*10), %f(31+i*10), %f(30+i*10);
ld.shared.f32 %f(33+i*10), [%r94+(8+i*8)];
ld.shared.f32 %f(34+i*10), [%r90+(8+i*8)];
fma.rn.f32 %f45, %f44, %f43, %f38;
for.size.1 i in range(0, 2, 1):
ld.shared.f32 %f(46+i*1), [%r(94+i*-4)+20];
for.size.2 i in range(0, 2, 1):
fma.rn.f32 %f(48+i*2), %f(47+i*-3), %f(46+i*3), %f(45+i*-3);
ld.shared.f32 %f(49+i*2), [%r94+(1584+i*4)];
for.size.3 i in range(0, 2, 1):
fma.rn.f32 %f(52+i*3), %f(47+i*7), %f(51+i*2), %f(50+i*-2);
ld.shared.f32 %f(53+i*3), [%r94+(24+i*4)];
ld.shared.f32 %f(54+i*3), [%r90+(24+i*4)];
for.size.2 i in range(0, 2, 1):
fma.rn.f32 %f(69+i*-10), %f(57+i*-3), %f(56+i*2), %f(55+i*-3);
ld.shared.f32 %f(58+i*2), [%r94+(1592+i*4)];
fma.rn.f32 %f68, %f57, %f60, %f59;
add.s32 %r102, %r102, 4;
setp.ne.s32 %p3, %r102, 196;
@%p3 bra $L__BB0_6;
add.s32 %r98, %r98, 1;
setp.lt.u32 %p4, %r98, 2;
@%p4 bra $L__BB0_1;
shl.b32 %r96, %r21, 6;
add.s32 %r97, %r96, %r1;
cvta.to.global.u64 %rd13, %rd5;
mul.wide.s32 %rd14, %r97, 4;
add.s64 %rd15, %rd13, %rd14;
ld.global.nc.f32 %f61, [%rd15];
add.f32 %f62, %f69, %f61;
mov.f32 %f63, 0f00000000;
max.f32 %f64, %f62, %f63;
cvta.to.global.u64 %rd16, %rd4;
add.s64 %rd17, %rd16, %rd14;
st.global.f32 [%rd17], %f64;
ld.global.nc.f32 %f65, [%rd15+4];
add.f32 %f66, %f68, %f65;
max.f32 %f67, %f66, %f63;
st.global.f32 [%rd17+4], %f67;
ret;
}
| extern "C" __global__ void default_function_kernel0(float* __restrict__ placeholder, float* __restrict__ placeholder1, float* __restrict__ T_relu, float* __restrict__ placeholder2) {
float T_dense[2];
__shared__ float placeholder_d_shared[392];
__shared__ float placeholder_shared[25088];
T_dense[0] = 0.000000e+00f;
T_dense[1] = 0.000000e+00f;
for (int k_outer_outer = 0; k_outer_outer < 2; ++k_outer_outer) {
__syncthreads();
int index0 = (int)threadIdx.x * 2;
int index1 = k_outer_outer * 392 + (int)threadIdx.x * 2;
placeholder_d_shared[index0] = placeholder[index1];
placeholder_d_shared[index0 + 1] = placeholder[index1 + 1];
placeholder_d_shared[index0 + 64] = placeholder[index1 + 64];
placeholder_d_shared[index0 + 65] = placeholder[index1 + 65];
placeholder_d_shared[index0 + 128] = placeholder[index1 + 128];
placeholder_d_shared[index0 + 129] = placeholder[index1 + 129];
placeholder_d_shared[index0 + 192] = placeholder[index1 + 192];
placeholder_d_shared[index0 + 193] = placeholder[index1 + 193];
placeholder_d_shared[index0 + 256] = placeholder[index1 + 256];
placeholder_d_shared[index0 + 257] = placeholder[index1 + 257];
placeholder_d_shared[index0 + 320] = placeholder[index1 + 320];
placeholder_d_shared[index0 + 321] = placeholder[index1 + 321];
if ((int)threadIdx.x < 4) {
placeholder_d_shared[index0 + 384] = placeholder[index1 + 384];
}
if ((int)threadIdx.x < 4) {
placeholder_d_shared[index0 + 385] = placeholder[index1 + 385];
}
for (int ax0_ax1_fused_outer_outer = 0; ax0_ax1_fused_outer_outer < 98; ++ax0_ax1_fused_outer_outer) {
ulonglong4*(placeholder_shared + ax0_ax1_fused_outer_outer * 256 + (int)threadIdx.x * 8)[0] = ulonglong4*(placeholder1 + (int)blockIdx.x * 50176 + (ax0_ax1_fused_outer_outer * 256 + (int)threadIdx.x * 8) / 392 * 784 + k_outer_outer * 392 + (ax0_ax1_fused_outer_outer * 256 + (int)threadIdx.x * 8) % 392)[0];
}
__syncthreads();
for (int k_outer_inner = 0; k_outer_inner < 196; ++k_outer_inner) {
int index2 = k_outer_inner * 2;
int index3 = (int)threadIdx.x * 784 + k_outer_inner * 2;
T_dense[0] = T_dense[0] + placeholder_d_shared[index2] * placeholder_shared[index3];
T_dense[0] = T_dense[0] + (placeholder_d_shared[index2 + 1] * placeholder_shared[index3 + 1]);
T_dense[1] = T_dense[1] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 392]);
T_dense[1] = T_dense[1] + (placeholder_d_shared[index2 + 1] * placeholder_shared[index3 + 393]);
}
}
for (int ax1_inner = 0; ax1_inner < 2; ++ax1_inner) {
int index4 = (int)blockIdx.x * 64 + (int)threadIdx.x * 2 + ax1_inner;
int index5 = ax1_inner;
T_relu[index4] = max(T_dense[index5] + placeholder2[index4], 0.000000e+00f);
}
} |
[407f97a0b39db63cc31f3e635caadfd0,1,784,128,784,128,1,128] | tmpf9aka2kn | default_function_kernel0 | .visible .entry default_function_kernel0(
.param .u64 default_function_kernel0_param_0,
.param .u64 default_function_kernel0_param_1,
.param .u64 default_function_kernel0_param_2,
.param .u64 default_function_kernel0_param_3
)
{
.reg .pred %p<3>;
.reg .f32 %f<197>;
.reg .b32 %r<18>;
.reg .b64 %rd<23>;
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E20placeholder_d_shared[8];
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E18placeholder_shared[1024];
ld.param.u64 %rd9, [default_function_kernel0_param_0];
ld.param.u64 %rd10, [default_function_kernel0_param_1];
ld.param.u64 %rd11, [default_function_kernel0_param_2];
ld.param.u64 %rd12, [default_function_kernel0_param_3];
cvta.to.global.u64 %rd13, %rd10;
cvta.to.global.u64 %rd1, %rd11;
cvta.to.global.u64 %rd2, %rd12;
mov.u32 %r1, %tid.x;
shl.b32 %r8, %r1, 2;
mov.u32 %r9, _ZZ24default_function_kernel0E20placeholder_d_shared;
add.s32 %r2, %r9, %r8;
shr.s32 %r10, %r1, 1;
and.b32 %r11, %r1, 1;
mov.u32 %r12, _ZZ24default_function_kernel0E18placeholder_shared;
add.s32 %r3, %r12, %r8;
shl.b32 %r13, %r1, 7;
add.s32 %r4, %r12, %r13;
cvta.to.global.u64 %rd14, %rd9;
mul.wide.s32 %rd15, %r1, 4;
add.s64 %rd22, %rd14, %rd15;
mul.lo.s32 %r14, %r10, 784;
or.b32 %r15, %r14, %r11;
mul.wide.s32 %rd16, %r15, 4;
add.s64 %rd17, %rd13, %rd16;
add.s64 %rd21, %rd17, 200704;
mov.u32 %r17, 0;
mov.f32 %f181, 0f00000000;
for.size.1 i in range(0, 15, 1):
mov.f32 %f(182+i*1), %f181;
$L__BB0_1:
bar.sync 0;
setp.gt.s32 %p1, %r1, 1;
@%p1 bra $L__BB0_3;
ld.global.nc.f32 %f49, [%rd22];
st.shared.f32 [%r2], %f49;
$L__BB0_3:
for.size.2 i in range(0, 32, 1):
ld.global.nc.f32 %f(50+i*1), [%rd21+(-200704+i*12544)];
st.shared.f32 [%r3+(0+i*32)], %f(50+i*1);
bar.sync 0;
for.size.33 i in range(0, 2, 1):
ld.shared.f32 %f(82+i*33), [%r4+(0+i*4)];
ld.shared.f32 %f(83+i*33), [_ZZ24default_function_kernel0E20placeholder_d_shared+(0+i*4)];
fma.rn.f32 %f(84+i*112), %f(83+i*33), %f(82+i*33), %f(196+i*-112);
ld.shared.f32 %f(85+i*32), [%r4+(8+i*4)];
fma.rn.f32 %f(86+i*109), %f(83+i*33), %f(85+i*32), %f(195+i*-109);
ld.shared.f32 %f(87+i*31), [%r4+(16+i*4)];
fma.rn.f32 %f(88+i*106), %f(83+i*33), %f(87+i*31), %f(194+i*-106);
ld.shared.f32 %f(89+i*30), [%r4+(24+i*4)];
fma.rn.f32 %f(90+i*103), %f(83+i*33), %f(89+i*30), %f(193+i*-103);
ld.shared.f32 %f(91+i*29), [%r4+(32+i*4)];
fma.rn.f32 %f(92+i*100), %f(83+i*33), %f(91+i*29), %f(192+i*-100);
ld.shared.f32 %f(93+i*28), [%r4+(40+i*4)];
fma.rn.f32 %f(94+i*97), %f(83+i*33), %f(93+i*28), %f(191+i*-97);
ld.shared.f32 %f(95+i*27), [%r4+(48+i*4)];
fma.rn.f32 %f(96+i*94), %f(83+i*33), %f(95+i*27), %f(190+i*-94);
ld.shared.f32 %f(97+i*26), [%r4+(56+i*4)];
fma.rn.f32 %f(98+i*91), %f(83+i*33), %f(97+i*26), %f(189+i*-91);
ld.shared.f32 %f(99+i*25), [%r4+(64+i*4)];
fma.rn.f32 %f(100+i*88), %f(83+i*33), %f(99+i*25), %f(188+i*-88);
ld.shared.f32 %f(101+i*24), [%r4+(72+i*4)];
fma.rn.f32 %f(102+i*85), %f(83+i*33), %f(101+i*24), %f(187+i*-85);
ld.shared.f32 %f(103+i*23), [%r4+(80+i*4)];
fma.rn.f32 %f(104+i*82), %f(83+i*33), %f(103+i*23), %f(186+i*-82);
ld.shared.f32 %f(105+i*22), [%r4+(88+i*4)];
fma.rn.f32 %f(106+i*79), %f(83+i*33), %f(105+i*22), %f(185+i*-79);
ld.shared.f32 %f(107+i*21), [%r4+(96+i*4)];
fma.rn.f32 %f(108+i*76), %f(83+i*33), %f(107+i*21), %f(184+i*-76);
ld.shared.f32 %f(109+i*20), [%r4+(104+i*4)];
fma.rn.f32 %f(110+i*73), %f(83+i*33), %f(109+i*20), %f(183+i*-73);
ld.shared.f32 %f(111+i*19), [%r4+(112+i*4)];
fma.rn.f32 %f(112+i*70), %f(83+i*33), %f(111+i*19), %f(182+i*-70);
ld.shared.f32 %f(113+i*18), [%r4+(120+i*4)];
fma.rn.f32 %f(114+i*67), %f(83+i*33), %f(113+i*18), %f(181+i*-67);
for.size.1 i in range(0, 2, 1):
add.s64 %rd(22+i*-1), %rd(22+i*-1), 8;
add.s32 %r17, %r17, 1;
setp.ne.s32 %p2, %r17, 392;
@%p2 bra $L__BB0_1;
shl.b32 %r16, %r1, 4;
mul.wide.s32 %rd18, %r16, 4;
add.s64 %rd19, %rd2, %rd18;
ld.global.nc.f32 %f132, [%rd19];
add.f32 %f133, %f196, %f132;
mov.f32 %f134, 0f00000000;
max.f32 %f135, %f133, %f134;
add.s64 %rd20, %rd1, %rd18;
for.size.4 i in range(0, 15, 1):
st.global.f32 [%rd20+(0+i*4)], %f(135+i*3);
ld.global.nc.f32 %f(136+i*3), [%rd19+(4+i*4)];
add.f32 %f(137+i*3), %f(195+i*-1), %f(136+i*3);
max.f32 %f(138+i*3), %f(137+i*3), %f134;
st.global.f32 [%rd20+60], %f180;
ret;
}
| extern "C" __global__ void default_function_kernel0(float* __restrict__ placeholder, float* __restrict__ placeholder1, float* __restrict__ T_relu, float* __restrict__ placeholder2) {
float T_dense[16];
__shared__ float placeholder_d_shared[2];
__shared__ float placeholder_shared[256];
for (int j_inner_init = 0; j_inner_init < 16; ++j_inner_init) {
int index0 = j_inner_init;
T_dense[index0] = 0.000000e+00f;
}
for (int k_outer_outer = 0; k_outer_outer < 392; ++k_outer_outer) {
__syncthreads();
if ((int)threadIdx.x < 2) {
int index1 = (int)threadIdx.x;
int index2 = k_outer_outer * 2 + (int)threadIdx.x;
placeholder_d_shared[index1] = placeholder[index2];
}
for (int ax0_ax1_fused_outer_outer = 0; ax0_ax1_fused_outer_outer < 32; ++ax0_ax1_fused_outer_outer) {
int index3 = ax0_ax1_fused_outer_outer * 8 + (int)threadIdx.x;
int index4 = ax0_ax1_fused_outer_outer * 3136 + ((int)threadIdx.x >> 1) * 784 + k_outer_outer * 2 + ((int)threadIdx.x & 1);
placeholder_shared[index3] = placeholder1[index4];
}
__syncthreads();
for (int k_outer_inner = 0; k_outer_inner < 2; ++k_outer_inner) {
for (int j_inner = 0; j_inner < 16; ++j_inner) {
int index5 = j_inner;
int index6 = k_outer_inner;
int index7 = (int)threadIdx.x * 32 + j_inner * 2 + k_outer_inner;
T_dense[index5] = T_dense[index5] + placeholder_d_shared[index6] * placeholder_shared[index7];
}
}
}
for (int ax1_inner = 0; ax1_inner < 16; ++ax1_inner) {
int index8 = (int)threadIdx.x * 16 + ax1_inner;
int index9 = ax1_inner;
T_relu[index8] = max(T_dense[index9] + placeholder2[index8], 0.000000e+00f);
}
} |
[407f97a0b39db63cc31f3e635caadfd0,1,784,128,784,128,1,128] | tmpjaqc7av6 | default_function_kernel0 | .visible .entry default_function_kernel0(
.param .u64 default_function_kernel0_param_0,
.param .u64 default_function_kernel0_param_1,
.param .u64 default_function_kernel0_param_2,
.param .u64 default_function_kernel0_param_3
)
{
.reg .pred %p<5>;
.reg .f32 %f<243>;
.reg .b32 %r<127>;
.reg .b64 %rd<30>;
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E20placeholder_d_shared[112];
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E18placeholder_shared[7168];
ld.param.u64 %rd5, [default_function_kernel0_param_0];
ld.param.u64 %rd6, [default_function_kernel0_param_1];
ld.param.u64 %rd7, [default_function_kernel0_param_2];
ld.param.u64 %rd8, [default_function_kernel0_param_3];
cvta.to.global.u64 %rd1, %rd6;
cvta.to.global.u64 %rd2, %rd5;
cvta.to.global.u64 %rd3, %rd7;
cvta.to.global.u64 %rd4, %rd8;
mov.u32 %r1, %tid.x;
shl.b32 %r26, %r1, 2;
mov.u32 %r27, _ZZ24default_function_kernel0E20placeholder_d_shared;
add.s32 %r2, %r27, %r26;
mov.u32 %r3, %ctaid.x;
for.size.1 i in range(0, 2, 1):
mul.lo.s32 %r(4+i*1), %r(3+i*-2), (50176+i*-50148);
add.s32 %r6, %r1, 28;
mov.u32 %r28, _ZZ24default_function_kernel0E18placeholder_shared;
add.s32 %r7, %r28, %r26;
mov.u32 %r121, 0;
mov.f32 %f227, 0f00000000;
for.size.1 i in range(0, 15, 1):
mov.f32 %f(228+i*1), %f227;
$L__BB0_1:
bar.sync 0;
mul.lo.s32 %r30, %r121, 28;
add.s32 %r31, %r30, %r1;
mul.wide.s32 %rd9, %r31, 4;
add.s64 %rd10, %rd2, %rd9;
for.size.2 i in range(0, 4, 1):
ld.global.nc.f32 %f(97+i*1), [%rd10+(0+i*16)];
st.shared.f32 [%r2+(0+i*16)], %f(97+i*1);
ld.global.nc.f32 %f101, [%rd10+64];
mov.u32 %r124, 64;
for.size.2 i in range(0, 2, 1):
st.shared.f32 [%r2+(64+i*16)], %f(101+i*1);
ld.global.nc.f32 %f(102+i*1), [%rd10+(80+i*16)];
st.shared.f32 [%r2+96], %f103;
add.s32 %r9, %r30, %r4;
mov.u32 %r122, %r1;
mov.u32 %r123, %r6;
$L__BB0_2:
mul.hi.s32 %r32, %r122, -1840700269;
add.s32 %r33, %r32, %r122;
shr.u32 %r34, %r33, 31;
shr.s32 %r35, %r33, 4;
add.s32 %r36, %r35, %r34;
mul.lo.s32 %r37, %r36, 28;
sub.s32 %r38, %r122, %r37;
add.s32 %r39, %r9, %r38;
mad.lo.s32 %r40, %r36, 784, %r39;
mul.wide.s32 %rd11, %r40, 4;
add.s64 %rd12, %rd1, %rd11;
ld.global.nc.f32 %f104, [%rd12];
add.s32 %r41, %r7, %r124;
for.size.14 i in range(0, 4, 1):
st.shared.f32 [%r41+(-64+i*16)], %f(104+i*1);
add.s32 %r(42+i*10), %r122, (4+i*4);
mul.hi.s32 %r(43+i*10), %r(42+i*10), -1840700269;
add.s32 %r(44+i*10), %r(43+i*10), %r(42+i*10);
shr.u32 %r(45+i*10), %r(44+i*10), 31;
shr.s32 %r(46+i*10), %r(44+i*10), 4;
add.s32 %r(47+i*10), %r(46+i*10), %r(45+i*10);
mul.lo.s32 %r(48+i*10), %r(47+i*10), 28;
sub.s32 %r(49+i*10), %r(42+i*10), %r(48+i*10);
add.s32 %r(50+i*10), %r9, %r(49+i*10);
mad.lo.s32 %r(51+i*10), %r(47+i*10), 784, %r(50+i*10);
mul.wide.s32 %rd(13+i*2), %r(51+i*10), 4;
add.s64 %rd(14+i*2), %rd1, %rd(13+i*2);
ld.global.nc.f32 %f(105+i*1), [%rd(14+i*2)+0];
for.size.14 i in range(0, 2, 1):
st.shared.f32 [%r41+(0+i*16)], %f(108+i*1);
add.s32 %r(82+i*10), %r123, (-8+i*4);
mul.hi.s32 %r(83+i*10), %r(82+i*10), -1840700269;
add.s32 %r(84+i*10), %r(83+i*10), %r(82+i*10);
shr.u32 %r(85+i*10), %r(84+i*10), 31;
shr.s32 %r(86+i*10), %r(84+i*10), 4;
add.s32 %r(87+i*10), %r(86+i*10), %r(85+i*10);
mul.lo.s32 %r(88+i*10), %r(87+i*10), 28;
sub.s32 %r(89+i*10), %r(82+i*10), %r(88+i*10);
add.s32 %r(90+i*10), %r9, %r(89+i*10);
mad.lo.s32 %r(91+i*10), %r(87+i*10), 784, %r(90+i*10);
mul.wide.s32 %rd(21+i*2), %r(91+i*10), 4;
add.s64 %rd(22+i*2), %rd1, %rd(21+i*2);
ld.global.nc.f32 %f(109+i*1), [%rd(22+i*2)+0];
st.shared.f32 [%r41+32], %f110;
mul.hi.s32 %r102, %r123, -1840700269;
add.s32 %r103, %r102, %r123;
shr.u32 %r104, %r103, 31;
shr.s32 %r105, %r103, 4;
add.s32 %r106, %r105, %r104;
mul.lo.s32 %r107, %r106, 28;
sub.s32 %r108, %r123, %r107;
add.s32 %r109, %r9, %r108;
mad.lo.s32 %r110, %r106, 784, %r109;
mul.wide.s32 %rd25, %r110, 4;
add.s64 %rd26, %rd1, %rd25;
ld.global.nc.f32 %f111, [%rd26];
st.shared.f32 [%r41+48], %f111;
add.s32 %r124, %r124, 128;
setp.ne.s32 %p1, %r124, 7232;
for.size.1 i in range(0, 2, 1):
add.s32 %r(14+i*108), %r123, (32+i*-28);
mov.u32 %r123, %r14;
@%p1 bra $L__BB0_2;
bar.sync 0;
mov.u32 %r111, 0;
mov.u32 %r125, %r111;
$L__BB0_4:
mul.lo.s32 %r17, %r125, 7;
add.s32 %r18, %r17, %r5;
mov.u32 %r126, %r111;
bra.uni $L__BB0_5;
$L__BB0_6:
for.size.1 i in range(0, 2, 1):
ld.shared.f32 %f(129+i*1), [%r(21+i*-1)+4];
for.size.2 i in range(0, 2, 1):
fma.rn.f32 %f(242+i*-1), %f130, %f(129+i*2), %f(242+i*-1);
ld.shared.f32 %f(131+i*1), [%r21+(452+i*448)];
for.size.2 i in range(0, 13, 1):
fma.rn.f32 %f(240+i*-1), %f130, %f(132+i*1), %f(240+i*-1);
ld.shared.f32 %f(133+i*1), [%r21+(1348+i*448)];
fma.rn.f32 %f227, %f130, %f145, %f227;
add.s32 %r126, %r126, 2;
$L__BB0_5:
for.size.3 i in range(0, 2, 1):
add.s32 %r(113+i*3), %r(126+i*-108), %r(17+i*109);
shl.b32 %r(114+i*3), %r(113+i*3), 2;
add.s32 %r(20+i*1), %r(27+i*1), %r(114+i*3);
for.size.1 i in range(0, 2, 1):
ld.shared.f32 %f(112+i*1), [%r(21+i*-1)+0];
for.size.2 i in range(0, 2, 1):
fma.rn.f32 %f(242+i*-1), %f113, %f(112+i*2), %f(242+i*-1);
ld.shared.f32 %f(114+i*1), [%r21+(448+i*448)];
for.size.2 i in range(0, 13, 1):
fma.rn.f32 %f(240+i*-1), %f113, %f(115+i*1), %f(240+i*-1);
ld.shared.f32 %f(116+i*1), [%r21+(1344+i*448)];
fma.rn.f32 %f227, %f113, %f128, %f227;
setp.eq.s32 %p2, %r126, 6;
@%p2 bra $L__BB0_7;
bra.uni $L__BB0_6;
$L__BB0_7:
add.s32 %r125, %r125, 1;
setp.lt.u32 %p3, %r125, 4;
@%p3 bra $L__BB0_4;
add.s32 %r121, %r121, 1;
setp.lt.u32 %p4, %r121, 28;
@%p4 bra $L__BB0_1;
shl.b32 %r119, %r3, 6;
add.s32 %r120, %r119, %r1;
mul.wide.s32 %rd27, %r120, 4;
add.s64 %rd28, %rd4, %rd27;
ld.global.nc.f32 %f146, [%rd28];
add.f32 %f147, %f242, %f146;
mov.f32 %f148, 0f00000000;
max.f32 %f149, %f147, %f148;
add.s64 %rd29, %rd3, %rd27;
for.size.4 i in range(0, 15, 1):
st.global.f32 [%rd29+(0+i*16)], %f(149+i*3);
ld.global.nc.f32 %f(150+i*3), [%rd28+(16+i*16)];
add.f32 %f(151+i*3), %f(241+i*-1), %f(150+i*3);
max.f32 %f(152+i*3), %f(151+i*3), %f148;
st.global.f32 [%rd29+240], %f194;
ret;
}
| extern "C" __global__ void default_function_kernel0(float* __restrict__ placeholder, float* __restrict__ placeholder1, float* __restrict__ T_relu, float* __restrict__ placeholder2) {
float T_dense[16];
__shared__ float placeholder_d_shared[28];
__shared__ float placeholder_shared[1792];
T_dense[0] = 0.000000e+00f;
T_dense[1] = 0.000000e+00f;
T_dense[2] = 0.000000e+00f;
T_dense[3] = 0.000000e+00f;
T_dense[4] = 0.000000e+00f;
T_dense[5] = 0.000000e+00f;
T_dense[6] = 0.000000e+00f;
T_dense[7] = 0.000000e+00f;
T_dense[8] = 0.000000e+00f;
T_dense[9] = 0.000000e+00f;
T_dense[10] = 0.000000e+00f;
T_dense[11] = 0.000000e+00f;
T_dense[12] = 0.000000e+00f;
T_dense[13] = 0.000000e+00f;
T_dense[14] = 0.000000e+00f;
T_dense[15] = 0.000000e+00f;
for (int k_outer_outer = 0; k_outer_outer < 28; ++k_outer_outer) {
__syncthreads();
int index0 = (int)threadIdx.x;
int index1 = k_outer_outer * 28 + (int)threadIdx.x;
placeholder_d_shared[index0] = placeholder[index1];
placeholder_d_shared[index0 + 4] = placeholder[index1 + 4];
placeholder_d_shared[index0 + 8] = placeholder[index1 + 8];
placeholder_d_shared[index0 + 12] = placeholder[index1 + 12];
placeholder_d_shared[index0 + 16] = placeholder[index1 + 16];
placeholder_d_shared[index0 + 20] = placeholder[index1 + 20];
placeholder_d_shared[index0 + 24] = placeholder[index1 + 24];
for (int ax0_ax1_fused_outer_outer = 0; ax0_ax1_fused_outer_outer < 448; ++ax0_ax1_fused_outer_outer) {
int index2 = ax0_ax1_fused_outer_outer * 4 + (int)threadIdx.x;
int index3 = (int)blockIdx.x * 50176 + (ax0_ax1_fused_outer_outer * 4 + (int)threadIdx.x) / 28 * 784 + k_outer_outer * 28 + (ax0_ax1_fused_outer_outer * 4 + (int)threadIdx.x) % 28;
placeholder_shared[index2] = placeholder1[index3];
}
__syncthreads();
for (int k_outer_inner = 0; k_outer_inner < 4; ++k_outer_inner) {
for (int k_inner = 0; k_inner < 7; ++k_inner) {
int index4 = k_outer_inner * 7 + k_inner;
int index5 = (int)threadIdx.x * 28 + k_outer_inner * 7 + k_inner;
T_dense[0] = T_dense[0] + placeholder_d_shared[index4] * placeholder_shared[index5];
T_dense[1] = T_dense[1] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 112]);
T_dense[2] = T_dense[2] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 224]);
T_dense[3] = T_dense[3] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 336]);
T_dense[4] = T_dense[4] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 448]);
T_dense[5] = T_dense[5] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 560]);
T_dense[6] = T_dense[6] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 672]);
T_dense[7] = T_dense[7] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 784]);
T_dense[8] = T_dense[8] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 896]);
T_dense[9] = T_dense[9] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1008]);
T_dense[10] = T_dense[10] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1120]);
T_dense[11] = T_dense[11] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1232]);
T_dense[12] = T_dense[12] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1344]);
T_dense[13] = T_dense[13] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1456]);
T_dense[14] = T_dense[14] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1568]);
T_dense[15] = T_dense[15] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1680]);
}
}
}
for (int vthread_s = 0; vthread_s < 16; ++vthread_s) {
int index6 = (int)blockIdx.x * 64 + vthread_s * 4 + (int)threadIdx.x;
int index7 = vthread_s;
T_relu[index6] = max(T_dense[index7] + placeholder2[index6], 0.000000e+00f);
}
} |
[407f97a0b39db63cc31f3e635caadfd0,1,784,128,784,128,1,128] | tmp1xwgmvpj | default_function_kernel0 | .visible .entry default_function_kernel0(
.param .u64 default_function_kernel0_param_0,
.param .u64 default_function_kernel0_param_1,
.param .u64 default_function_kernel0_param_2,
.param .u64 default_function_kernel0_param_3
)
{
.reg .pred %p<5>;
.reg .f32 %f<659>;
.reg .b32 %r<48>;
.reg .b64 %rd<24>;
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E20placeholder_d_shared[8];
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E18placeholder_shared[1024];
ld.param.u64 %rd8, [default_function_kernel0_param_0];
ld.param.u64 %rd9, [default_function_kernel0_param_1];
cvta.to.global.u64 %rd12, %rd9;
cvta.to.global.u64 %rd2, %rd8;
mov.u32 %r1, %tid.x;
shl.b32 %r16, %r1, 3;
mov.u32 %r17, _ZZ24default_function_kernel0E20placeholder_d_shared;
add.s32 %r2, %r17, %r16;
mul.lo.s32 %r3, %r1, 784;
shl.b32 %r18, %r1, 4;
mov.u32 %r19, _ZZ24default_function_kernel0E18placeholder_shared;
add.s32 %r20, %r19, %r18;
add.s32 %r5, %r20, 64;
add.s64 %rd4, %rd12, 25088;
mul.lo.s32 %r6, %r1, 1568;
mov.u32 %r44, 0;
mov.f32 %f531, 0f00000000;
for.size.1 i in range(0, 63, 1):
mov.f32 %f(532+i*1), %f531;
$L__BB0_1:
bar.sync 0;
setp.lt.s32 %p1, %r1, 1;
@%p1 bra $L__BB0_2;
bra.uni $L__BB0_3;
$L__BB0_2:
shl.b32 %r21, %r44, 1;
add.s32 %r22, %r21, %r3;
mul.wide.s32 %rd13, %r22, 4;
add.s64 %rd14, %rd2, %rd13;
ld.global.nc.v2.u32 {%r23, %r24}, [%rd14];
st.shared.v2.u32 [%r2], {%r23, %r24};
$L__BB0_3:
shl.b32 %r28, %r44, 1;
add.s32 %r29, %r6, %r28;
mul.wide.s32 %rd15, %r29, 4;
add.s64 %rd23, %rd4, %rd15;
mov.u32 %r46, 64;
mov.u32 %r45, %r5;
$L__BB0_4:
for.size.2 i in range(0, 2, 1):
for.size.1 j in range(0, 2, 1):
ld.global.nc.f32 %f(257+i*2+j*1), [%rd23+(-21948+i*-3136+j*-4)];
for.size.5 i in range(0, 3, 1):
st.shared.v4.f32 [%r45+(-64+i*32)], {%f(260+i*4), %f(259+i*4), %f(258+i*4), %f(257+i*4)};
ld.global.nc.f32 %f(261+i*4), [%rd23+(-9404+i*12544)];
ld.global.nc.f32 %f(262+i*4), [%rd23+(-9408+i*12544)];
ld.global.nc.f32 %f(263+i*4), [%rd23+(-12540+i*12544)];
ld.global.nc.f32 %f(264+i*4), [%rd23+(-12544+i*12544)];
st.shared.v4.f32 [%r45+32], {%f272, %f271, %f270, %f269};
add.s32 %r45, %r45, 128;
add.s64 %rd23, %rd23, 50176;
add.s32 %r46, %r46, 128;
setp.ne.s32 %p2, %r46, 1088;
@%p2 bra $L__BB0_4;
bar.sync 0;
mov.u32 %r47, 0;
$L__BB0_6:
mov.u32 %r43, _ZZ24default_function_kernel0E20placeholder_d_shared;
mov.u32 %r42, _ZZ24default_function_kernel0E18placeholder_shared;
for.size.2 i in range(0, 2, 1):
shl.b32 %r(41+i*-8), %r(1+i*31), (7+i*-5);
add.s32 %r(32+i*3), %r(47+i*-5), %r(41+i*-8);
ld.shared.f32 %f273, [%r35];
shl.b32 %r36, %r47, 2;
add.s32 %r38, %r43, %r36;
ld.shared.f32 %f274, [%r38];
for.size.2 i in range(0, 2, 1):
fma.rn.f32 %f(594+i*-1), %f274, %f(273+i*2), %f(594+i*-1);
ld.shared.f32 %f(275+i*1), [%r35+(8+i*8)];
for.size.2 i in range(0, 61, 1):
fma.rn.f32 %f(592+i*-1), %f274, %f(276+i*1), %f(592+i*-1);
ld.shared.f32 %f(277+i*1), [%r35+(24+i*8)];
fma.rn.f32 %f531, %f274, %f337, %f531;
add.s32 %r39, %r47, 1;
mov.u32 %r47, 1;
setp.lt.u32 %p3, %r39, 2;
@%p3 bra $L__BB0_6;
add.s32 %r44, %r44, 1;
setp.lt.u32 %p4, %r44, 392;
@%p4 bra $L__BB0_1;
ld.param.u64 %rd22, [default_function_kernel0_param_2];
cvta.to.global.u64 %rd21, %rd22;
ld.param.u64 %rd20, [default_function_kernel0_param_3];
cvta.to.global.u64 %rd19, %rd20;
shl.b32 %r40, %r1, 6;
mul.wide.s32 %rd16, %r40, 4;
add.s64 %rd17, %rd19, %rd16;
ld.global.nc.f32 %f338, [%rd17];
add.f32 %f339, %f594, %f338;
mov.f32 %f340, 0f00000000;
max.f32 %f341, %f339, %f340;
add.s64 %rd18, %rd21, %rd16;
for.size.4 i in range(0, 63, 1):
st.global.f32 [%rd18+(0+i*4)], %f(341+i*3);
ld.global.nc.f32 %f(342+i*3), [%rd17+(4+i*4)];
add.f32 %f(343+i*3), %f(593+i*-1), %f(342+i*3);
max.f32 %f(344+i*3), %f(343+i*3), %f340;
st.global.f32 [%rd18+252], %f530;
ret;
}
| extern "C" __global__ void default_function_kernel0(float* __restrict__ placeholder, float* __restrict__ placeholder1, float* __restrict__ T_relu, float* __restrict__ placeholder2) {
float T_dense[64];
__shared__ float placeholder_d_shared[2];
__shared__ float placeholder_shared[256];
for (int j_inner_init = 0; j_inner_init < 64; ++j_inner_init) {
int index0 = j_inner_init;
T_dense[index0] = 0.000000e+00f;
}
for (int k_outer_outer = 0; k_outer_outer < 392; ++k_outer_outer) {
__syncthreads();
if ((int)threadIdx.x < 1) {
((float2*)(placeholder_d_shared + (int)threadIdx.x * 2))[0] = ((float2*)(placeholder + (int)threadIdx.x * 784 + k_outer_outer * 2))[0];
}
for (int ax0_ax1_fused_outer_outer = 0; ax0_ax1_fused_outer_outer < 32; ++ax0_ax1_fused_outer_outer) {
int4 _1;
int4 _2;
int4 _3;
int4 _4 = make_int4(ax0_ax1_fused_outer_outer * 3136 + (int)threadIdx.x * 1568, ax0_ax1_fused_outer_outer * 3136 + (int)threadIdx.x * 1568, ax0_ax1_fused_outer_outer * 3136 + (int)threadIdx.x * 1568, ax0_ax1_fused_outer_outer * 3136 + (int)threadIdx.x * 1568);
int4 _5;
int4 _6;
int4 _7 = make_int4(0+1*0, 0+1*1, 0+1*2, 0+1*3);
int4 _8 = make_int4(2, 2, 2, 2);
_6.x = _7.x%_8.x;
_6.y = _7.y%_8.y;
_6.z = _7.z%_8.z;
_6.w = _7.w%_8.w;
int4 _9;
int4 _10 = make_int4(0+1*0, 0+1*1, 0+1*2, 0+1*3);
int4 _11 = make_int4(2, 2, 2, 2);
_9.x = _10.x/_11.x;
_9.y = _10.y/_11.y;
_9.z = _10.z/_11.z;
_9.w = _10.w/_11.w;
int4 _12;
ushort4 _13;
ushort4 _14;
ushort4 _15;
int4 _16 = make_int4(2, 2, 2, 2);
int4 _17 = make_int4(0, 0, 0, 0);
_15.x = _16.x>=_17.x;
_15.y = _16.y>=_17.y;
_15.z = _16.z>=_17.z;
_15.w = _16.w>=_17.w;
ushort4 _18;
int4 _19 = make_int4(0, 0, 0, 0);
_18.x = _6.x>=_19.x;
_18.y = _6.y>=_19.y;
_18.z = _6.z>=_19.z;
_18.w = _6.w>=_19.w;
_14.x = _15.x&&_18.x;
_14.y = _15.y&&_18.y;
_14.z = _15.z&&_18.z;
_14.w = _15.w&&_18.w;
ushort4 _20;
ushort4 _21;
int4 _22 = make_int4(2, 2, 2, 2);
int4 _23 = make_int4(0, 0, 0, 0);
_21.x = _22.x<_23.x;
_21.y = _22.y<_23.y;
_21.z = _22.z<_23.z;
_21.w = _22.w<_23.w;
ushort4 _24;
int4 _25 = make_int4(0, 0, 0, 0);
_24.x = _6.x<=_25.x;
_24.y = _6.y<=_25.y;
_24.z = _6.z<=_25.z;
_24.w = _6.w<=_25.w;
_20.x = _21.x&&_24.x;
_20.y = _21.y&&_24.y;
_20.z = _21.z&&_24.z;
_20.w = _21.w&&_24.w;
_13.x = _14.x||_20.x;
_13.y = _14.y||_20.y;
_13.z = _14.z||_20.z;
_13.w = _14.w||_20.w;
int4 _26;
int4 _27 = make_int4(1, 1, 1, 1);
_26.x = _9.x-_27.x;
_26.y = _9.y-_27.y;
_26.z = _9.z-_27.z;
_26.w = _9.w-_27.w;
_12.x = bool(_13.x)?_9.x:_26.x;
_12.y = bool(_13.y)?_9.y:_26.y;
_12.z = bool(_13.z)?_9.z:_26.z;
_12.w = bool(_13.w)?_9.w:_26.w;
int4 _28 = make_int4(784, 784, 784, 784);
_5.x = _12.x*_28.x;
_5.y = _12.y*_28.y;
_5.z = _12.z*_28.z;
_5.w = _12.w*_28.w;
_3.x = _4.x+_5.x;
_3.y = _4.y+_5.y;
_3.z = _4.z+_5.z;
_3.w = _4.w+_5.w;
int4 _29 = make_int4(k_outer_outer * 2, k_outer_outer * 2, k_outer_outer * 2, k_outer_outer * 2);
_2.x = _3.x+_29.x;
_2.y = _3.y+_29.y;
_2.z = _3.z+_29.z;
_2.w = _3.w+_29.w;
int4 _30;
int4 _31 = make_int4(0+1*0, 0+1*1, 0+1*2, 0+1*3);
int4 _32 = make_int4(2, 2, 2, 2);
_30.x = _31.x%_32.x;
_30.y = _31.y%_32.y;
_30.z = _31.z%_32.z;
_30.w = _31.w%_32.w;
int4 _33;
ushort4 _34;
ushort4 _35;
ushort4 _36;
int4 _37 = make_int4(2, 2, 2, 2);
int4 _38 = make_int4(0, 0, 0, 0);
_36.x = _37.x>=_38.x;
_36.y = _37.y>=_38.y;
_36.z = _37.z>=_38.z;
_36.w = _37.w>=_38.w;
ushort4 _39;
int4 _40 = make_int4(0, 0, 0, 0);
_39.x = _30.x>=_40.x;
_39.y = _30.y>=_40.y;
_39.z = _30.z>=_40.z;
_39.w = _30.w>=_40.w;
_35.x = _36.x&&_39.x;
_35.y = _36.y&&_39.y;
_35.z = _36.z&&_39.z;
_35.w = _36.w&&_39.w;
ushort4 _41;
ushort4 _42;
int4 _43 = make_int4(2, 2, 2, 2);
int4 _44 = make_int4(0, 0, 0, 0);
_42.x = _43.x<_44.x;
_42.y = _43.y<_44.y;
_42.z = _43.z<_44.z;
_42.w = _43.w<_44.w;
ushort4 _45;
int4 _46 = make_int4(0, 0, 0, 0);
_45.x = _30.x<=_46.x;
_45.y = _30.y<=_46.y;
_45.z = _30.z<=_46.z;
_45.w = _30.w<=_46.w;
_41.x = _42.x&&_45.x;
_41.y = _42.y&&_45.y;
_41.z = _42.z&&_45.z;
_41.w = _42.w&&_45.w;
_34.x = _35.x||_41.x;
_34.y = _35.y||_41.y;
_34.z = _35.z||_41.z;
_34.w = _35.w||_41.w;
int4 _47;
int4 _48 = make_int4(2, 2, 2, 2);
_47.x = _30.x+_48.x;
_47.y = _30.y+_48.y;
_47.z = _30.z+_48.z;
_47.w = _30.w+_48.w;
_33.x = bool(_34.x)?_30.x:_47.x;
_33.y = bool(_34.y)?_30.y:_47.y;
_33.z = bool(_34.z)?_30.z:_47.z;
_33.w = bool(_34.w)?_30.w:_47.w;
_1.x = _2.x+_33.x;
_1.y = _2.y+_33.y;
_1.z = _2.z+_33.z;
_1.w = _2.w+_33.w;
int index1 = _1.x;
int index2 = _1.y;
int index3 = _1.z;
int index4 = _1.w;
((float4*)(placeholder_shared + ax0_ax1_fused_outer_outer * 8 + (int)threadIdx.x * 4))[0] = make_float4(placeholder1[index1],placeholder1[index2],placeholder1[index3],placeholder1[index4]);
}
__syncthreads();
for (int k_outer_inner = 0; k_outer_inner < 2; ++k_outer_inner) {
for (int j_inner = 0; j_inner < 64; ++j_inner) {
int index5 = j_inner;
int index6 = k_outer_inner;
int index7 = (int)threadIdx.x * 128 + j_inner * 2 + k_outer_inner;
T_dense[index5] = T_dense[index5] + placeholder_d_shared[index6] * placeholder_shared[index7];
}
}
}
for (int ax1_inner = 0; ax1_inner < 64; ++ax1_inner) {
int index8 = (int)threadIdx.x * 64 + ax1_inner;
int index9 = ax1_inner;
T_relu[index8] = max(T_dense[index9] + placeholder2[index8], 0.000000e+00f);
}
} |
[407f97a0b39db63cc31f3e635caadfd0,1,784,128,784,128,1,128] | tmpddyc20a0 | default_function_kernel0 | .visible .entry default_function_kernel0(
.param .u64 default_function_kernel0_param_0,
.param .u64 default_function_kernel0_param_1,
.param .u64 default_function_kernel0_param_2,
.param .u64 default_function_kernel0_param_3
)
{
.reg .pred %p<4>;
.reg .f32 %f<70>;
.reg .b32 %r<39>;
.reg .b64 %rd<26>;
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E20placeholder_d_shared[16];
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E18placeholder_shared[1024];
ld.param.u64 %rd11, [default_function_kernel0_param_0];
ld.param.u64 %rd12, [default_function_kernel0_param_1];
ld.param.u64 %rd13, [default_function_kernel0_param_2];
ld.param.u64 %rd10, [default_function_kernel0_param_3];
cvta.to.global.u64 %rd14, %rd12;
cvta.to.global.u64 %rd1, %rd13;
mov.u32 %r1, %tid.x;
for.size.1 i in range(0, 2, 1):
shl.b32 %r(10+i*1), %r1, (2+i*2);
mov.u32 %r12, _ZZ24default_function_kernel0E20placeholder_d_shared;
add.s32 %r2, %r12, %r11;
shr.s32 %r13, %r1, 2;
and.b32 %r14, %r1, 3;
mov.u32 %r15, _ZZ24default_function_kernel0E18placeholder_shared;
for.size.1 i in range(0, 2, 1):
add.s32 %r(3+i*1), %r15, %r(10+i*1);
mul.lo.s32 %r16, %r1, 784;
cvta.to.global.u64 %rd15, %rd11;
mul.wide.s32 %rd16, %r16, 4;
add.s64 %rd17, %rd15, %rd16;
add.s64 %rd25, %rd17, 16;
mov.u32 %r5, %ctaid.x;
mul.lo.s32 %r17, %r13, 784;
mad.lo.s32 %r18, %r5, 50176, %r17;
or.b32 %r19, %r18, %r14;
mul.wide.s32 %rd18, %r19, 4;
add.s64 %rd19, %rd14, %rd18;
add.s64 %rd24, %rd19, 100352;
mov.f32 %f68, 0f00000000;
mov.u32 %r38, 0;
mov.f32 %f69, %f68;
$L__BB0_1:
bar.sync 0;
add.s64 %rd6, %rd25, -16;
setp.gt.s32 %p1, %r1, 0;
@%p1 bra $L__BB0_3;
ld.global.nc.v4.u32 {%r20, %r21, %r22, %r23}, [%rd6];
st.shared.v4.u32 [%r2], {%r20, %r21, %r22, %r23};
$L__BB0_3:
add.s64 %rd7, %rd24, -100352;
for.size.2 i in range(0, 8, 1):
ld.global.nc.f32 %f(9+i*1), [%rd24+(-100352+i*25088)];
st.shared.f32 [%r3+(0+i*128)], %f(9+i*1);
bar.sync 0;
for.size.5 i in range(0, 2, 1):
ld.shared.f32 %f(17+i*5), [%r4+(0+i*4)];
ld.shared.f32 %f(18+i*5), [_ZZ24default_function_kernel0E20placeholder_d_shared+(0+i*4)];
fma.rn.f32 %f(19+i*5), %f(18+i*5), %f(17+i*5), %f(68+i*-49);
ld.shared.f32 %f(20+i*5), [%r4+(512+i*4)];
fma.rn.f32 %f(21+i*5), %f(18+i*5), %f(20+i*5), %f(69+i*-48);
for.size.5 i in range(0, 2, 1):
ld.shared.f32 %f(27+i*5), [%r4+(8+i*4)];
ld.shared.f32 %f(28+i*5), [_ZZ24default_function_kernel0E20placeholder_d_shared+(8+i*4)];
fma.rn.f32 %f(29+i*-26), %f(28+i*5), %f(27+i*5), %f(24+i*5);
ld.shared.f32 %f(30+i*4), [%r4+(520+i*4)];
fma.rn.f32 %f(31+i*-27), %f(28+i*5), %f(30+i*4), %f(26+i*5);
bar.sync 0;
@%p1 bra $L__BB0_5;
ld.global.nc.v4.u32 {%r28, %r29, %r30, %r31}, [%rd6+16];
st.shared.v4.u32 [%r2], {%r28, %r29, %r30, %r31};
$L__BB0_5:
for.size.2 i in range(0, 8, 1):
ld.global.nc.f32 %f(35+i*1), [%rd7+(16+i*25088)];
st.shared.f32 [%r3+(0+i*128)], %f(35+i*1);
bar.sync 0;
for.size.5 i in range(0, 2, 1):
ld.shared.f32 %f(43+i*5), [%r4+(0+i*4)];
ld.shared.f32 %f(44+i*5), [_ZZ24default_function_kernel0E20placeholder_d_shared+(0+i*4)];
fma.rn.f32 %f(45+i*5), %f(44+i*5), %f(43+i*5), %f(3+i*42);
ld.shared.f32 %f(46+i*5), [%r4+(512+i*4)];
fma.rn.f32 %f(47+i*5), %f(44+i*5), %f(46+i*5), %f(4+i*43);
for.size.5 i in range(0, 2, 1):
ld.shared.f32 %f(53+i*5), [%r4+(8+i*4)];
ld.shared.f32 %f(54+i*5), [_ZZ24default_function_kernel0E20placeholder_d_shared+(8+i*4)];
fma.rn.f32 %f(55+i*13), %f(54+i*5), %f(53+i*5), %f(50+i*5);
ld.shared.f32 %f(56+i*4), [%r4+(520+i*4)];
fma.rn.f32 %f(57+i*12), %f(54+i*5), %f(56+i*4), %f(52+i*5);
for.size.1 i in range(0, 2, 1):
add.s64 %rd(25+i*-1), %rd(25+i*-1), 32;
add.s32 %r38, %r38, 2;
setp.ne.s32 %p3, %r38, 196;
@%p3 bra $L__BB0_1;
cvta.to.global.u64 %rd20, %rd10;
shl.b32 %r36, %r5, 6;
add.s32 %r37, %r36, %r1;
mul.wide.s32 %rd21, %r37, 4;
add.s64 %rd22, %rd20, %rd21;
ld.global.nc.f32 %f61, [%rd22];
add.f32 %f62, %f68, %f61;
mov.f32 %f63, 0f00000000;
max.f32 %f64, %f62, %f63;
add.s64 %rd23, %rd1, %rd21;
st.global.f32 [%rd23], %f64;
ld.global.nc.f32 %f65, [%rd22+128];
add.f32 %f66, %f69, %f65;
max.f32 %f67, %f66, %f63;
st.global.f32 [%rd23+128], %f67;
ret;
}
| extern "C" __global__ void default_function_kernel0(float* __restrict__ placeholder, float* __restrict__ placeholder1, float* __restrict__ T_relu, float* __restrict__ placeholder2) {
float T_dense[2];
__shared__ float placeholder_d_shared[4];
__shared__ float placeholder_shared[256];
T_dense[0] = 0.000000e+00f;
T_dense[1] = 0.000000e+00f;
for (int k_outer_outer = 0; k_outer_outer < 196; ++k_outer_outer) {
__syncthreads();
if ((int)threadIdx.x < 1) {
((float4*)(placeholder_d_shared + (int)threadIdx.x * 4))[0] = ((float4*)(placeholder + (int)threadIdx.x * 784 + k_outer_outer * 4))[0];
}
int index0 = (int)threadIdx.x;
int index1 = (int)blockIdx.x * 50176 + ((int)threadIdx.x >> 2) * 784 + k_outer_outer * 4 + ((int)threadIdx.x & 3);
placeholder_shared[index0] = placeholder1[index1];
placeholder_shared[index0 + 32] = placeholder1[index1 + 6272];
placeholder_shared[index0 + 64] = placeholder1[index1 + 12544];
placeholder_shared[index0 + 96] = placeholder1[index1 + 18816];
placeholder_shared[index0 + 128] = placeholder1[index1 + 25088];
placeholder_shared[index0 + 160] = placeholder1[index1 + 31360];
placeholder_shared[index0 + 192] = placeholder1[index1 + 37632];
placeholder_shared[index0 + 224] = placeholder1[index1 + 43904];
__syncthreads();
T_dense[0] = T_dense[0] + placeholder_d_shared[0] * placeholder_shared[index0 * 4];
int index2 = (int)threadIdx.x * 4 + 128;
T_dense[1] = T_dense[1] + placeholder_d_shared[0] * placeholder_shared[index2];
int index3 = (int)threadIdx.x * 4 + 1;
T_dense[0] = T_dense[0] + placeholder_d_shared[1] * placeholder_shared[index3];
int index4 = (int)threadIdx.x * 4 + 129;
T_dense[1] = T_dense[1] + placeholder_d_shared[1] * placeholder_shared[index4];
int index5 = (int)threadIdx.x * 4 + 2;
T_dense[0] = T_dense[0] + placeholder_d_shared[2] * placeholder_shared[index5];
int index6 = (int)threadIdx.x * 4 + 130;
T_dense[1] = T_dense[1] + placeholder_d_shared[2] * placeholder_shared[index6];
int index7 = (int)threadIdx.x * 4 + 3;
T_dense[0] = T_dense[0] + placeholder_d_shared[3] * placeholder_shared[index7];
int index8 = (int)threadIdx.x * 4 + 131;
T_dense[1] = T_dense[1] + placeholder_d_shared[3] * placeholder_shared[index8];
}
int index9 = (int)blockIdx.x * 64 + (int)threadIdx.x;
T_relu[index9] = max(T_dense[0] + placeholder2[index9], 0.000000e+00f);
T_relu[index9 + 32] = max(T_dense[1] + placeholder2[index9 + 32], 0.000000e+00f);
} |
[407f97a0b39db63cc31f3e635caadfd0,1,784,128,784,128,1,128] | tmpt255kjz8 | default_function_kernel0 | .visible .entry default_function_kernel0(
.param .u64 default_function_kernel0_param_0,
.param .u64 default_function_kernel0_param_1,
.param .u64 default_function_kernel0_param_2,
.param .u64 default_function_kernel0_param_3
)
{
.reg .pred %p<6>;
.reg .f32 %f<779>;
.reg .b32 %r<137>;
.reg .b64 %rd<32>;
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E20placeholder_d_shared[1568];
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E18placeholder_shared[200704];
ld.param.u64 %rd1, [default_function_kernel0_param_0];
ld.param.u64 %rd2, [default_function_kernel0_param_1];
mov.u32 %r14, %tid.x;
mul.lo.s32 %r1, %r14, 6272;
mov.u32 %r132, 0;
mov.f32 %f715, 0f00000000;
cvta.to.global.u64 %rd5, %rd1;
cvta.to.global.u64 %rd8, %rd2;
for.size.1 i in range(0, 63, 1):
mov.f32 %f(716+i*1), %f715;
$L__BB0_1:
mov.u32 %r133, 0;
bar.sync 0;
$L__BB0_2:
mul.lo.s32 %r129, %r132, 392;
for.size.1 i in range(0, 2, 1):
shl.b32 %r(17+i*1), %r(14+i*119), (1+i*1);
add.s32 %r19, %r18, %r17;
shl.b32 %r20, %r19, 2;
mov.u32 %r21, _ZZ24default_function_kernel0E20placeholder_d_shared;
for.size.1 i in range(0, 2, 1):
add.s32 %r(22+i*1), %r(21+i*108), %r(20+i*-3);
add.s32 %r24, %r23, %r18;
mul.wide.s32 %rd6, %r24, 4;
add.s64 %rd7, %rd5, %rd6;
for.size.2 i in range(0, 2, 1):
ld.global.nc.v2.u32 {%r(25+i*4), %r(26+i*4)}, [%rd7+(0+i*16)];
st.shared.v2.u32 [%r22+(0+i*16)], {%r(25+i*4), %r(26+i*4)};
add.s32 %r133, %r133, 2;
setp.ne.s32 %p1, %r133, 98;
@%p1 bra $L__BB0_2;
mov.u32 %r134, 0;
$L__BB0_4:
mul.lo.s32 %r125, %r132, 392;
shl.b32 %r35, %r134, 1;
add.s32 %r36, %r35, %r14;
mul.hi.s32 %r37, %r36, 1402438301;
shr.u32 %r38, %r37, 31;
shr.s32 %r39, %r37, 7;
add.s32 %r40, %r39, %r38;
mul.lo.s32 %r41, %r40, 392;
sub.s32 %r42, %r36, %r41;
add.s32 %r43, %r42, %r125;
mad.lo.s32 %r44, %r40, 784, %r43;
mul.wide.s32 %rd9, %r44, 4;
add.s64 %rd10, %rd8, %rd9;
ld.global.nc.f32 %f321, [%rd10];
shl.b32 %r45, %r36, 2;
mov.u32 %r46, _ZZ24default_function_kernel0E18placeholder_shared;
add.s32 %r47, %r46, %r45;
for.size.13 i in range(0, 7, 1):
st.shared.f32 [%r47+(0+i*8)], %f(321+i*1);
add.s32 %r(48+i*9), %r36, (2+i*2);
mul.hi.s32 %r(49+i*9), %r(48+i*9), 1402438301;
shr.u32 %r(50+i*9), %r(49+i*9), 31;
shr.s32 %r(51+i*9), %r(49+i*9), 7;
add.s32 %r(52+i*9), %r(51+i*9), %r(50+i*9);
mul.lo.s32 %r(53+i*9), %r(52+i*9), 392;
sub.s32 %r(54+i*9), %r(48+i*9), %r(53+i*9);
add.s32 %r(55+i*9), %r(54+i*9), %r125;
mad.lo.s32 %r(56+i*9), %r(52+i*9), 784, %r(55+i*9);
mul.wide.s32 %rd(11+i*2), %r(56+i*9), 4;
add.s64 %rd(12+i*2), %rd8, %rd(11+i*2);
ld.global.nc.f32 %f(322+i*1), [%rd(12+i*2)+0];
st.shared.f32 [%r47+56], %f328;
add.s32 %r134, %r134, 8;
setp.ne.s32 %p2, %r134, 25088;
@%p2 bra $L__BB0_4;
bar.sync 0;
mov.u32 %r135, 0;
$L__BB0_6:
mov.u32 %r136, 0;
$L__BB0_7:
shl.b32 %r131, %r135, 1;
mov.u32 %r127, _ZZ24default_function_kernel0E18placeholder_shared;
mov.u32 %r126, _ZZ24default_function_kernel0E20placeholder_d_shared;
for.size.3 i in range(0, 2, 1):
add.s32 %r(114+i*4), %r(136+i*-22), %r(131+i*-130);
shl.b32 %r(115+i*4), %r(114+i*4), 2;
add.s32 %r(117+i*4), %r(126+i*1), %r(115+i*4);
for.size.1 i in range(0, 2, 1):
ld.shared.f32 %f(329+i*1), [%r(121+i*-4)+0];
for.size.2 i in range(0, 2, 1):
fma.rn.f32 %f(778+i*-16), %f330, %f(329+i*2), %f(778+i*-16);
ld.shared.f32 %f(331+i*1), [%r121+(50176+i*50176)];
for.size.2 i in range(0, 2, 1):
fma.rn.f32 %f(746+i*-16), %f330, %f(332+i*1), %f(746+i*-16);
ld.shared.f32 %f(333+i*1), [%r121+(150528+i*-148960)];
for.size.2 i in range(0, 3, 1):
fma.rn.f32 %f(777+i*-16), %f330, %f(334+i*1), %f(777+i*-16);
ld.shared.f32 %f(335+i*1), [%r121+(51744+i*50176)];
for.size.6 i in range(0, 14, 1):
for.size.2 j in range(0, 2, 1):
fma.rn.f32 %f(729+i*-1+j*47), %f330, %f(337+i*4+j*1), %f(729+i*-1+j*47);
ld.shared.f32 %f(338+i*4+j*1), [%r121+(3136+i*1568+j*50176)];
for.size.2 j in range(0, 2, 1):
fma.rn.f32 %f(760+i*-1+j*-16), %f330, %f(339+i*4+j*1), %f(760+i*-1+j*-16);
ld.shared.f32 %f(340+i*4+j*1), [%r121+(103488+i*1568+j*50176)];
fma.rn.f32 %f715, %f330, %f393, %f715;
add.s32 %r122, %r136, 1;
mov.u32 %r136, 1;
setp.lt.u32 %p3, %r122, 2;
@%p3 bra $L__BB0_7;
add.s32 %r135, %r135, 1;
setp.lt.u32 %p4, %r135, 196;
@%p4 bra $L__BB0_6;
add.s32 %r132, %r132, 1;
setp.lt.u32 %p5, %r132, 2;
@%p5 bra $L__BB0_1;
ld.param.u64 %rd31, [default_function_kernel0_param_2];
ld.param.u64 %rd30, [default_function_kernel0_param_3];
shl.b32 %r124, %r14, 4;
cvta.to.global.u64 %rd25, %rd30;
mul.wide.s32 %rd26, %r124, 4;
add.s64 %rd27, %rd25, %rd26;
ld.global.nc.f32 %f394, [%rd27];
add.f32 %f395, %f778, %f394;
mov.f32 %f396, 0f00000000;
max.f32 %f397, %f395, %f396;
cvta.to.global.u64 %rd28, %rd31;
add.s64 %rd29, %rd28, %rd26;
for.size.4 i in range(0, 3, 1):
st.global.f32 [%rd29+(0+i*128)], %f(397+i*3);
ld.global.nc.f32 %f(398+i*3), [%rd27+(128+i*128)];
add.f32 %f(399+i*3), %f(762+i*-16), %f(398+i*3);
max.f32 %f(400+i*3), %f(399+i*3), %f396;
for.size.10 i in range(0, 15, 1):
for.size.4 j in range(0, 2, 1):
st.global.f32 [%rd29+(384+i*4+j*-380)], %f(406+i*12+j*3);
ld.global.nc.f32 %f(407+i*12+j*3), [%rd27+(4+i*4+j*128)];
add.f32 %f(408+i*12+j*3), %f(777+i*-1+j*-16), %f(407+i*12+j*3);
max.f32 %f(409+i*12+j*3), %f(408+i*12+j*3), %f396;
for.size.4 j in range(0, 2, 1):
st.global.f32 [%rd29+(132+i*4+j*128)], %f(412+i*12+j*3);
ld.global.nc.f32 %f(413+i*12+j*3), [%rd27+(260+i*4+j*128)];
add.f32 %f(414+i*12+j*3), %f(745+i*-1+j*-16), %f(413+i*12+j*3);
max.f32 %f(415+i*12+j*3), %f(414+i*12+j*3), %f396;
st.global.f32 [%rd29+444], %f586;
ret;
}
| extern "C" __global__ void default_function_kernel0(float* __restrict__ placeholder, float* __restrict__ placeholder1, float* __restrict__ T_relu, float* __restrict__ placeholder2) {
float T_dense[64];
__shared__ float placeholder_d_shared[392];
__shared__ float placeholder_shared[50176];
for (int j_inner_init = 0; j_inner_init < 16; ++j_inner_init) {
int index0 = j_inner_init;
T_dense[index0] = 0.000000e+00f;
T_dense[index0 + 16] = 0.000000e+00f;
T_dense[index0 + 32] = 0.000000e+00f;
T_dense[index0 + 48] = 0.000000e+00f;
}
for (int k_outer_outer = 0; k_outer_outer < 2; ++k_outer_outer) {
__syncthreads();
for (int ax0_ax1_fused_outer_outer = 0; ax0_ax1_fused_outer_outer < 98; ++ax0_ax1_fused_outer_outer) {
((float2*)(placeholder_d_shared + ax0_ax1_fused_outer_outer * 4 + (int)threadIdx.x * 2))[0] = ((float2*)(placeholder + k_outer_outer * 392 + ax0_ax1_fused_outer_outer * 4 + (int)threadIdx.x * 2))[0];
}
for (int ax0_ax1_fused_outer_outer1 = 0; ax0_ax1_fused_outer_outer1 < 25088; ++ax0_ax1_fused_outer_outer1) {
int index1 = ax0_ax1_fused_outer_outer1 * 2 + (int)threadIdx.x;
int index2 = (ax0_ax1_fused_outer_outer1 * 2 + (int)threadIdx.x) / 392 * 784 + k_outer_outer * 392 + (ax0_ax1_fused_outer_outer1 * 2 + (int)threadIdx.x) % 392;
placeholder_shared[index1] = placeholder1[index2];
}
__syncthreads();
for (int k_outer_inner = 0; k_outer_inner < 196; ++k_outer_inner) {
for (int k_inner = 0; k_inner < 2; ++k_inner) {
for (int j_inner = 0; j_inner < 16; ++j_inner) {
int index3 = j_inner;
int index4 = k_outer_inner * 2 + k_inner;
int index5 = (int)threadIdx.x * 6272 + j_inner * 392 + k_outer_inner * 2 + k_inner;
T_dense[index3] = T_dense[index3] + placeholder_d_shared[index4] * placeholder_shared[index5];
T_dense[index3 + 16] = T_dense[index3 + 16] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 12544]);
T_dense[index3 + 32] = T_dense[index3 + 32] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 25088]);
T_dense[index3 + 48] = T_dense[index3 + 48] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 37632]);
}
}
}
}
for (int ax1_inner = 0; ax1_inner < 16; ++ax1_inner) {
int index6 = (int)threadIdx.x * 16 + ax1_inner;
int index7 = ax1_inner;
T_relu[index6] = max(T_dense[index7] + placeholder2[index6], 0.000000e+00f);
T_relu[index6 + 32] = max(T_dense[index7 + 16] + placeholder2[index6 + 32], 0.000000e+00f);
T_relu[index6 + 64] = max(T_dense[index7 + 32] + placeholder2[index6 + 64], 0.000000e+00f);
T_relu[index6 + 96] = max(T_dense[index7 + 48] + placeholder2[index6 + 96], 0.000000e+00f);
}
} |
[407f97a0b39db63cc31f3e635caadfd0,1,784,128,784,128,1,128] | tmphyw2eb5o | default_function_kernel0 | .visible .entry default_function_kernel0(
.param .u64 default_function_kernel0_param_0,
.param .u64 default_function_kernel0_param_1,
.param .u64 default_function_kernel0_param_2,
.param .u64 default_function_kernel0_param_3
)
{
.reg .pred %p<5>;
.reg .f32 %f<807>;
.reg .b32 %r<127>;
.reg .b64 %rd<32>;
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E20placeholder_d_shared[224];
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E18placeholder_shared[28672];
ld.param.u64 %rd1, [default_function_kernel0_param_0];
ld.param.u64 %rd2, [default_function_kernel0_param_1];
mov.u32 %r13, %tid.x;
mul.lo.s32 %r1, %r13, 3584;
mov.u32 %r123, 0;
mov.f32 %f743, 0f00000000;
cvta.to.global.u64 %rd5, %rd1;
shl.b32 %r17, %r13, 2;
mov.u32 %r18, _ZZ24default_function_kernel0E20placeholder_d_shared;
add.s32 %r19, %r18, %r17;
cvta.to.global.u64 %rd8, %rd2;
for.size.1 i in range(0, 63, 1):
mov.f32 %f(744+i*1), %f743;
$L__BB0_1:
mov.u32 %r124, 0;
bar.sync 0;
mul.lo.s32 %r3, %r123, 56;
add.s32 %r16, %r3, %r13;
mul.wide.s32 %rd6, %r16, 4;
add.s64 %rd7, %rd5, %rd6;
for.size.2 i in range(0, 28, 1):
ld.global.nc.f32 %f(321+i*1), [%rd7+(0+i*8)];
st.shared.f32 [%r19+(0+i*8)], %f(321+i*1);
$L__BB0_2:
mul.lo.s32 %r117, %r123, 56;
shl.b32 %r21, %r124, 1;
add.s32 %r22, %r21, %r13;
mul.hi.s32 %r23, %r22, -1840700269;
add.s32 %r24, %r23, %r22;
shr.u32 %r25, %r24, 31;
shr.s32 %r26, %r24, 5;
add.s32 %r27, %r26, %r25;
mul.lo.s32 %r28, %r27, 56;
sub.s32 %r29, %r22, %r28;
add.s32 %r30, %r29, %r117;
mad.lo.s32 %r31, %r27, 784, %r30;
mul.wide.s32 %rd9, %r31, 4;
add.s64 %rd10, %rd8, %rd9;
ld.global.nc.f32 %f349, [%rd10];
shl.b32 %r32, %r22, 2;
mov.u32 %r33, _ZZ24default_function_kernel0E18placeholder_shared;
add.s32 %r34, %r33, %r32;
for.size.14 i in range(0, 7, 1):
st.shared.f32 [%r34+(0+i*8)], %f(349+i*1);
add.s32 %r(35+i*10), %r22, (2+i*2);
mul.hi.s32 %r(36+i*10), %r(35+i*10), -1840700269;
add.s32 %r(37+i*10), %r(36+i*10), %r(35+i*10);
shr.u32 %r(38+i*10), %r(37+i*10), 31;
shr.s32 %r(39+i*10), %r(37+i*10), 5;
add.s32 %r(40+i*10), %r(39+i*10), %r(38+i*10);
mul.lo.s32 %r(41+i*10), %r(40+i*10), 56;
sub.s32 %r(42+i*10), %r(35+i*10), %r(41+i*10);
add.s32 %r(43+i*10), %r(42+i*10), %r117;
mad.lo.s32 %r(44+i*10), %r(40+i*10), 784, %r(43+i*10);
mul.wide.s32 %rd(11+i*2), %r(44+i*10), 4;
add.s64 %rd(12+i*2), %rd8, %rd(11+i*2);
ld.global.nc.f32 %f(350+i*1), [%rd(12+i*2)+0];
st.shared.f32 [%r34+56], %f356;
add.s32 %r124, %r124, 8;
setp.ne.s32 %p1, %r124, 3584;
@%p1 bra $L__BB0_2;
bar.sync 0;
mov.u32 %r125, 0;
$L__BB0_4:
mov.u32 %r126, 0;
$L__BB0_5:
mul.lo.s32 %r121, %r125, 14;
mov.u32 %r119, _ZZ24default_function_kernel0E18placeholder_shared;
mov.u32 %r118, _ZZ24default_function_kernel0E20placeholder_d_shared;
for.size.3 i in range(0, 2, 1):
add.s32 %r(107+i*4), %r(126+i*-19), %r(121+i*-120);
shl.b32 %r(108+i*4), %r(107+i*4), 2;
add.s32 %r(110+i*4), %r(118+i*1), %r(108+i*4);
for.size.1 i in range(0, 2, 1):
ld.shared.f32 %f(357+i*1), [%r(114+i*-4)+0];
for.size.2 i in range(0, 2, 1):
fma.rn.f32 %f(806+i*-1), %f358, %f(357+i*2), %f(806+i*-1);
ld.shared.f32 %f(359+i*1), [%r114+(224+i*224)];
for.size.2 i in range(0, 61, 1):
fma.rn.f32 %f(804+i*-1), %f358, %f(360+i*1), %f(804+i*-1);
ld.shared.f32 %f(361+i*1), [%r114+(672+i*224)];
fma.rn.f32 %f743, %f358, %f421, %f743;
add.s32 %r126, %r126, 1;
setp.ne.s32 %p2, %r126, 14;
@%p2 bra $L__BB0_5;
add.s32 %r125, %r125, 1;
setp.lt.u32 %p3, %r125, 4;
@%p3 bra $L__BB0_4;
add.s32 %r123, %r123, 1;
setp.lt.u32 %p4, %r123, 14;
@%p4 bra $L__BB0_1;
ld.param.u64 %rd31, [default_function_kernel0_param_2];
ld.param.u64 %rd30, [default_function_kernel0_param_3];
shl.b32 %r116, %r13, 6;
cvta.to.global.u64 %rd25, %rd30;
mul.wide.s32 %rd26, %r116, 4;
add.s64 %rd27, %rd25, %rd26;
ld.global.nc.f32 %f422, [%rd27];
add.f32 %f423, %f806, %f422;
mov.f32 %f424, 0f00000000;
max.f32 %f425, %f423, %f424;
cvta.to.global.u64 %rd28, %rd31;
add.s64 %rd29, %rd28, %rd26;
for.size.4 i in range(0, 63, 1):
st.global.f32 [%rd29+(0+i*4)], %f(425+i*3);
ld.global.nc.f32 %f(426+i*3), [%rd27+(4+i*4)];
add.f32 %f(427+i*3), %f(805+i*-1), %f(426+i*3);
max.f32 %f(428+i*3), %f(427+i*3), %f424;
st.global.f32 [%rd29+252], %f614;
ret;
}
| extern "C" __global__ void default_function_kernel0(float* __restrict__ placeholder, float* __restrict__ placeholder1, float* __restrict__ T_relu, float* __restrict__ placeholder2) {
float T_dense[64];
__shared__ float placeholder_d_shared[56];
__shared__ float placeholder_shared[7168];
T_dense[0] = 0.000000e+00f;
T_dense[1] = 0.000000e+00f;
T_dense[2] = 0.000000e+00f;
T_dense[3] = 0.000000e+00f;
T_dense[4] = 0.000000e+00f;
T_dense[5] = 0.000000e+00f;
T_dense[6] = 0.000000e+00f;
T_dense[7] = 0.000000e+00f;
T_dense[8] = 0.000000e+00f;
T_dense[9] = 0.000000e+00f;
T_dense[10] = 0.000000e+00f;
T_dense[11] = 0.000000e+00f;
T_dense[12] = 0.000000e+00f;
T_dense[13] = 0.000000e+00f;
T_dense[14] = 0.000000e+00f;
T_dense[15] = 0.000000e+00f;
T_dense[16] = 0.000000e+00f;
T_dense[17] = 0.000000e+00f;
T_dense[18] = 0.000000e+00f;
T_dense[19] = 0.000000e+00f;
T_dense[20] = 0.000000e+00f;
T_dense[21] = 0.000000e+00f;
T_dense[22] = 0.000000e+00f;
T_dense[23] = 0.000000e+00f;
T_dense[24] = 0.000000e+00f;
T_dense[25] = 0.000000e+00f;
T_dense[26] = 0.000000e+00f;
T_dense[27] = 0.000000e+00f;
T_dense[28] = 0.000000e+00f;
T_dense[29] = 0.000000e+00f;
T_dense[30] = 0.000000e+00f;
T_dense[31] = 0.000000e+00f;
T_dense[32] = 0.000000e+00f;
T_dense[33] = 0.000000e+00f;
T_dense[34] = 0.000000e+00f;
T_dense[35] = 0.000000e+00f;
T_dense[36] = 0.000000e+00f;
T_dense[37] = 0.000000e+00f;
T_dense[38] = 0.000000e+00f;
T_dense[39] = 0.000000e+00f;
T_dense[40] = 0.000000e+00f;
T_dense[41] = 0.000000e+00f;
T_dense[42] = 0.000000e+00f;
T_dense[43] = 0.000000e+00f;
T_dense[44] = 0.000000e+00f;
T_dense[45] = 0.000000e+00f;
T_dense[46] = 0.000000e+00f;
T_dense[47] = 0.000000e+00f;
T_dense[48] = 0.000000e+00f;
T_dense[49] = 0.000000e+00f;
T_dense[50] = 0.000000e+00f;
T_dense[51] = 0.000000e+00f;
T_dense[52] = 0.000000e+00f;
T_dense[53] = 0.000000e+00f;
T_dense[54] = 0.000000e+00f;
T_dense[55] = 0.000000e+00f;
T_dense[56] = 0.000000e+00f;
T_dense[57] = 0.000000e+00f;
T_dense[58] = 0.000000e+00f;
T_dense[59] = 0.000000e+00f;
T_dense[60] = 0.000000e+00f;
T_dense[61] = 0.000000e+00f;
T_dense[62] = 0.000000e+00f;
T_dense[63] = 0.000000e+00f;
for (int k_outer_outer = 0; k_outer_outer < 14; ++k_outer_outer) {
__syncthreads();
int index0 = (int)threadIdx.x;
int index1 = k_outer_outer * 56 + (int)threadIdx.x;
placeholder_d_shared[index0] = placeholder[index1];
placeholder_d_shared[index0 + 2] = placeholder[index1 + 2];
placeholder_d_shared[index0 + 4] = placeholder[index1 + 4];
placeholder_d_shared[index0 + 6] = placeholder[index1 + 6];
placeholder_d_shared[index0 + 8] = placeholder[index1 + 8];
placeholder_d_shared[index0 + 10] = placeholder[index1 + 10];
placeholder_d_shared[index0 + 12] = placeholder[index1 + 12];
placeholder_d_shared[index0 + 14] = placeholder[index1 + 14];
placeholder_d_shared[index0 + 16] = placeholder[index1 + 16];
placeholder_d_shared[index0 + 18] = placeholder[index1 + 18];
placeholder_d_shared[index0 + 20] = placeholder[index1 + 20];
placeholder_d_shared[index0 + 22] = placeholder[index1 + 22];
placeholder_d_shared[index0 + 24] = placeholder[index1 + 24];
placeholder_d_shared[index0 + 26] = placeholder[index1 + 26];
placeholder_d_shared[index0 + 28] = placeholder[index1 + 28];
placeholder_d_shared[index0 + 30] = placeholder[index1 + 30];
placeholder_d_shared[index0 + 32] = placeholder[index1 + 32];
placeholder_d_shared[index0 + 34] = placeholder[index1 + 34];
placeholder_d_shared[index0 + 36] = placeholder[index1 + 36];
placeholder_d_shared[index0 + 38] = placeholder[index1 + 38];
placeholder_d_shared[index0 + 40] = placeholder[index1 + 40];
placeholder_d_shared[index0 + 42] = placeholder[index1 + 42];
placeholder_d_shared[index0 + 44] = placeholder[index1 + 44];
placeholder_d_shared[index0 + 46] = placeholder[index1 + 46];
placeholder_d_shared[index0 + 48] = placeholder[index1 + 48];
placeholder_d_shared[index0 + 50] = placeholder[index1 + 50];
placeholder_d_shared[index0 + 52] = placeholder[index1 + 52];
placeholder_d_shared[index0 + 54] = placeholder[index1 + 54];
for (int ax0_ax1_fused_outer_outer = 0; ax0_ax1_fused_outer_outer < 3584; ++ax0_ax1_fused_outer_outer) {
int index2 = ax0_ax1_fused_outer_outer * 2 + (int)threadIdx.x;
int index3 = (ax0_ax1_fused_outer_outer * 2 + (int)threadIdx.x) / 56 * 784 + k_outer_outer * 56 + (ax0_ax1_fused_outer_outer * 2 + (int)threadIdx.x) % 56;
placeholder_shared[index2] = placeholder1[index3];
}
__syncthreads();
for (int k_outer_inner = 0; k_outer_inner < 4; ++k_outer_inner) {
for (int k_inner = 0; k_inner < 14; ++k_inner) {
int index4 = k_outer_inner * 14 + k_inner;
int index5 = (int)threadIdx.x * 3584 + k_outer_inner * 14 + k_inner;
T_dense[0] = T_dense[0] + placeholder_d_shared[index4] * placeholder_shared[index5];
T_dense[1] = T_dense[1] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 56]);
T_dense[2] = T_dense[2] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 112]);
T_dense[3] = T_dense[3] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 168]);
T_dense[4] = T_dense[4] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 224]);
T_dense[5] = T_dense[5] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 280]);
T_dense[6] = T_dense[6] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 336]);
T_dense[7] = T_dense[7] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 392]);
T_dense[8] = T_dense[8] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 448]);
T_dense[9] = T_dense[9] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 504]);
T_dense[10] = T_dense[10] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 560]);
T_dense[11] = T_dense[11] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 616]);
T_dense[12] = T_dense[12] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 672]);
T_dense[13] = T_dense[13] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 728]);
T_dense[14] = T_dense[14] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 784]);
T_dense[15] = T_dense[15] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 840]);
T_dense[16] = T_dense[16] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 896]);
T_dense[17] = T_dense[17] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 952]);
T_dense[18] = T_dense[18] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1008]);
T_dense[19] = T_dense[19] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1064]);
T_dense[20] = T_dense[20] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1120]);
T_dense[21] = T_dense[21] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1176]);
T_dense[22] = T_dense[22] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1232]);
T_dense[23] = T_dense[23] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1288]);
T_dense[24] = T_dense[24] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1344]);
T_dense[25] = T_dense[25] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1400]);
T_dense[26] = T_dense[26] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1456]);
T_dense[27] = T_dense[27] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1512]);
T_dense[28] = T_dense[28] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1568]);
T_dense[29] = T_dense[29] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1624]);
T_dense[30] = T_dense[30] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1680]);
T_dense[31] = T_dense[31] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1736]);
T_dense[32] = T_dense[32] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1792]);
T_dense[33] = T_dense[33] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1848]);
T_dense[34] = T_dense[34] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1904]);
T_dense[35] = T_dense[35] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 1960]);
T_dense[36] = T_dense[36] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 2016]);
T_dense[37] = T_dense[37] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 2072]);
T_dense[38] = T_dense[38] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 2128]);
T_dense[39] = T_dense[39] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 2184]);
T_dense[40] = T_dense[40] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 2240]);
T_dense[41] = T_dense[41] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 2296]);
T_dense[42] = T_dense[42] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 2352]);
T_dense[43] = T_dense[43] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 2408]);
T_dense[44] = T_dense[44] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 2464]);
T_dense[45] = T_dense[45] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 2520]);
T_dense[46] = T_dense[46] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 2576]);
T_dense[47] = T_dense[47] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 2632]);
T_dense[48] = T_dense[48] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 2688]);
T_dense[49] = T_dense[49] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 2744]);
T_dense[50] = T_dense[50] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 2800]);
T_dense[51] = T_dense[51] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 2856]);
T_dense[52] = T_dense[52] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 2912]);
T_dense[53] = T_dense[53] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 2968]);
T_dense[54] = T_dense[54] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 3024]);
T_dense[55] = T_dense[55] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 3080]);
T_dense[56] = T_dense[56] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 3136]);
T_dense[57] = T_dense[57] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 3192]);
T_dense[58] = T_dense[58] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 3248]);
T_dense[59] = T_dense[59] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 3304]);
T_dense[60] = T_dense[60] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 3360]);
T_dense[61] = T_dense[61] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 3416]);
T_dense[62] = T_dense[62] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 3472]);
T_dense[63] = T_dense[63] + (placeholder_d_shared[index4] * placeholder_shared[index5 + 3528]);
}
}
}
for (int ax1_inner = 0; ax1_inner < 64; ++ax1_inner) {
int index6 = (int)threadIdx.x * 64 + ax1_inner;
int index7 = ax1_inner;
T_relu[index6] = max(T_dense[index7] + placeholder2[index6], 0.000000e+00f);
}
} |
[407f97a0b39db63cc31f3e635caadfd0,1,784,128,784,128,1,128] | tmpsbahwun6 | default_function_kernel0 | .visible .entry default_function_kernel0(
.param .u64 default_function_kernel0_param_0,
.param .u64 default_function_kernel0_param_1,
.param .u64 default_function_kernel0_param_2,
.param .u64 default_function_kernel0_param_3
)
{
.reg .pred %p<5>;
.reg .f32 %f<667>;
.reg .b32 %r<47>;
.reg .b64 %rd<27>;
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E20placeholder_d_shared[3136];
// demoted variable
.shared .align 4 .b8 _ZZ24default_function_kernel0E18placeholder_shared[401408];
ld.param.u64 %rd11, [default_function_kernel0_param_0];
ld.param.u64 %rd8, [default_function_kernel0_param_1];
cvta.to.global.u64 %rd12, %rd11;
mov.u32 %r17, %tid.x;
shl.b32 %r18, %r17, 2;
mov.u32 %r19, _ZZ24default_function_kernel0E20placeholder_d_shared;
add.s32 %r20, %r19, %r18;
add.s32 %r41, %r20, 32;
cvt.s64.s32 %rd1, %r17;
mul.wide.s32 %rd13, %r17, 4;
add.s64 %rd14, %rd12, %rd13;
add.s64 %rd25, %rd14, 32;
mov.u32 %r42, 0;
$L__BB0_1:
for.size.2 i in range(0, 8, 1):
ld.global.nc.f32 %f(193+i*1), [%rd25+(-32+i*8)];
st.shared.f32 [%r41+(-32+i*8)], %f(193+i*1);
add.s32 %r41, %r41, 64;
add.s64 %rd25, %rd25, 64;
add.s32 %r42, %r42, 8;
setp.ne.s32 %p1, %r42, 392;
@%p1 bra $L__BB0_1;
cvt.u32.u64 %r22, %rd1;
shl.b32 %r23, %r22, 2;
mov.u32 %r24, _ZZ24default_function_kernel0E18placeholder_shared;
add.s32 %r25, %r24, %r23;
add.s32 %r43, %r25, 64;
cvta.to.global.u64 %rd15, %rd8;
shl.b64 %rd16, %rd1, 2;
add.s64 %rd17, %rd15, %rd16;
add.s64 %rd26, %rd17, 64;
mov.u32 %r44, 0;
$L__BB0_3:
for.size.2 i in range(0, 16, 1):
ld.global.nc.f32 %f(201+i*1), [%rd26+(-64+i*8)];
st.shared.f32 [%r43+(-64+i*8)], %f(201+i*1);
add.s32 %r43, %r43, 128;
add.s64 %rd26, %rd26, 128;
add.s32 %r44, %r44, 16;
setp.ne.s32 %p2, %r44, 50176;
@%p2 bra $L__BB0_3;
bar.sync 0;
mov.u32 %r26, 0;
mov.f32 %f603, 0f00000000;
for.size.1 i in range(0, 63, 1):
mov.f32 %f(604+i*1), %f603;
mov.u32 %r45, %r26;
$L__BB0_5:
mov.u32 %r46, %r26;
$L__BB0_6:
mul.lo.s32 %r40, %r45, 56;
mov.u32 %r39, _ZZ24default_function_kernel0E20placeholder_d_shared;
add.s32 %r28, %r46, %r40;
shl.b32 %r29, %r28, 2;
add.s32 %r31, %r39, %r29;
mad.lo.s32 %r33, %r17, 12544, %r28;
shl.b32 %r34, %r33, 2;
add.s32 %r36, %r24, %r34;
for.size.1 i in range(0, 2, 1):
ld.shared.f32 %f(281+i*1), [%r(36+i*-5)+0];
for.size.2 i in range(0, 2, 1):
fma.rn.f32 %f(666+i*-1), %f282, %f(281+i*2), %f(666+i*-1);
ld.shared.f32 %f(283+i*1), [%r36+(100352+i*100352)];
for.size.2 i in range(0, 2, 1):
fma.rn.f32 %f(664+i*-1), %f282, %f(284+i*1), %f(664+i*-1);
ld.shared.f32 %f(285+i*1), [%r36+(301056+i*-297920)];
for.size.2 i in range(0, 3, 1):
fma.rn.f32 %f(662+i*-1), %f282, %f(286+i*1), %f(662+i*-1);
ld.shared.f32 %f(287+i*1), [%r36+(103488+i*100352)];
for.size.3 i in range(0, 14, 1):
for.size.2 j in range(0, 4, 1):
fma.rn.f32 %f(659+i*-4+j*-1), %f282, %f(289+i*4+j*1), %f(659+i*-4+j*-1);
ld.shared.f32 %f(290+i*4+j*1), [%r36+(6272+i*3136+j*100352)];
fma.rn.f32 %f603, %f282, %f345, %f603;
add.s32 %r46, %r46, 1;
setp.ne.s32 %p3, %r46, 56;
@%p3 bra $L__BB0_6;
add.s32 %r45, %r45, 1;
setp.lt.u32 %p4, %r45, 14;
@%p4 bra $L__BB0_5;
ld.param.u64 %rd24, [default_function_kernel0_param_2];
ld.param.u64 %rd23, [default_function_kernel0_param_3];
shl.b32 %r38, %r17, 4;
cvta.to.global.u64 %rd18, %rd23;
mul.wide.s32 %rd19, %r38, 4;
add.s64 %rd20, %rd18, %rd19;
ld.global.nc.f32 %f346, [%rd20];
add.f32 %f347, %f666, %f346;
mov.f32 %f348, 0f00000000;
max.f32 %f349, %f347, %f348;
cvta.to.global.u64 %rd21, %rd24;
add.s64 %rd22, %rd21, %rd19;
for.size.4 i in range(0, 3, 1):
st.global.f32 [%rd22+(0+i*128)], %f(349+i*3);
ld.global.nc.f32 %f(350+i*3), [%rd20+(128+i*128)];
add.f32 %f(351+i*3), %f(665+i*-1), %f(350+i*3);
max.f32 %f(352+i*3), %f(351+i*3), %f348;
for.size.10 i in range(0, 15, 1):
for.size.4 j in range(0, 2, 1):
st.global.f32 [%rd22+(384+i*4+j*-380)], %f(358+i*12+j*3);
ld.global.nc.f32 %f(359+i*12+j*3), [%rd20+(4+i*4+j*128)];
add.f32 %f(360+i*12+j*3), %f(662+i*-4+j*-1), %f(359+i*12+j*3);
max.f32 %f(361+i*12+j*3), %f(360+i*12+j*3), %f348;
for.size.4 j in range(0, 2, 1):
st.global.f32 [%rd22+(132+i*4+j*128)], %f(364+i*12+j*3);
ld.global.nc.f32 %f(365+i*12+j*3), [%rd20+(260+i*4+j*128)];
add.f32 %f(366+i*12+j*3), %f(660+i*-4+j*-1), %f(365+i*12+j*3);
max.f32 %f(367+i*12+j*3), %f(366+i*12+j*3), %f348;
st.global.f32 [%rd22+444], %f538;
ret;
}
| extern "C" __global__ void default_function_kernel0(float* __restrict__ placeholder, float* __restrict__ placeholder1, float* __restrict__ T_relu, float* __restrict__ placeholder2) {
float T_dense[64];
__shared__ float placeholder_d_shared[784];
__shared__ float placeholder_shared[100352];
T_dense[0] = 0.000000e+00f;
T_dense[16] = 0.000000e+00f;
T_dense[32] = 0.000000e+00f;
T_dense[48] = 0.000000e+00f;
T_dense[1] = 0.000000e+00f;
T_dense[17] = 0.000000e+00f;
T_dense[33] = 0.000000e+00f;
T_dense[49] = 0.000000e+00f;
T_dense[2] = 0.000000e+00f;
T_dense[18] = 0.000000e+00f;
T_dense[34] = 0.000000e+00f;
T_dense[50] = 0.000000e+00f;
T_dense[3] = 0.000000e+00f;
T_dense[19] = 0.000000e+00f;
T_dense[35] = 0.000000e+00f;
T_dense[51] = 0.000000e+00f;
T_dense[4] = 0.000000e+00f;
T_dense[20] = 0.000000e+00f;
T_dense[36] = 0.000000e+00f;
T_dense[52] = 0.000000e+00f;
T_dense[5] = 0.000000e+00f;
T_dense[21] = 0.000000e+00f;
T_dense[37] = 0.000000e+00f;
T_dense[53] = 0.000000e+00f;
T_dense[6] = 0.000000e+00f;
T_dense[22] = 0.000000e+00f;
T_dense[38] = 0.000000e+00f;
T_dense[54] = 0.000000e+00f;
T_dense[7] = 0.000000e+00f;
T_dense[23] = 0.000000e+00f;
T_dense[39] = 0.000000e+00f;
T_dense[55] = 0.000000e+00f;
T_dense[8] = 0.000000e+00f;
T_dense[24] = 0.000000e+00f;
T_dense[40] = 0.000000e+00f;
T_dense[56] = 0.000000e+00f;
T_dense[9] = 0.000000e+00f;
T_dense[25] = 0.000000e+00f;
T_dense[41] = 0.000000e+00f;
T_dense[57] = 0.000000e+00f;
T_dense[10] = 0.000000e+00f;
T_dense[26] = 0.000000e+00f;
T_dense[42] = 0.000000e+00f;
T_dense[58] = 0.000000e+00f;
T_dense[11] = 0.000000e+00f;
T_dense[27] = 0.000000e+00f;
T_dense[43] = 0.000000e+00f;
T_dense[59] = 0.000000e+00f;
T_dense[12] = 0.000000e+00f;
T_dense[28] = 0.000000e+00f;
T_dense[44] = 0.000000e+00f;
T_dense[60] = 0.000000e+00f;
T_dense[13] = 0.000000e+00f;
T_dense[29] = 0.000000e+00f;
T_dense[45] = 0.000000e+00f;
T_dense[61] = 0.000000e+00f;
T_dense[14] = 0.000000e+00f;
T_dense[30] = 0.000000e+00f;
T_dense[46] = 0.000000e+00f;
T_dense[62] = 0.000000e+00f;
T_dense[15] = 0.000000e+00f;
T_dense[31] = 0.000000e+00f;
T_dense[47] = 0.000000e+00f;
T_dense[63] = 0.000000e+00f;
for (int ax0_ax1_fused_outer_outer = 0; ax0_ax1_fused_outer_outer < 392; ++ax0_ax1_fused_outer_outer) {
int index0 = ax0_ax1_fused_outer_outer * 2 + (int)threadIdx.x;
placeholder_d_shared[index0] = placeholder[index0];
}
for (int ax0_ax1_fused_outer_outer1 = 0; ax0_ax1_fused_outer_outer1 < 50176; ++ax0_ax1_fused_outer_outer1) {
int index1 = ax0_ax1_fused_outer_outer1 * 2 + (int)threadIdx.x;
placeholder_shared[index1] = placeholder1[index1];
}
__syncthreads();
for (int k_outer_inner = 0; k_outer_inner < 14; ++k_outer_inner) {
for (int k_inner = 0; k_inner < 56; ++k_inner) {
int index2 = k_outer_inner * 56 + k_inner;
int index3 = (int)threadIdx.x * 12544 + k_outer_inner * 56 + k_inner;
T_dense[0] = T_dense[0] + placeholder_d_shared[index2] * placeholder_shared[index3];
T_dense[16] = T_dense[16] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 25088]);
T_dense[32] = T_dense[32] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 50176]);
T_dense[48] = T_dense[48] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 75264]);
T_dense[1] = T_dense[1] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 784]);
T_dense[17] = T_dense[17] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 25872]);
T_dense[33] = T_dense[33] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 50960]);
T_dense[49] = T_dense[49] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 76048]);
T_dense[2] = T_dense[2] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 1568]);
T_dense[18] = T_dense[18] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 26656]);
T_dense[34] = T_dense[34] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 51744]);
T_dense[50] = T_dense[50] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 76832]);
T_dense[3] = T_dense[3] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 2352]);
T_dense[19] = T_dense[19] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 27440]);
T_dense[35] = T_dense[35] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 52528]);
T_dense[51] = T_dense[51] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 77616]);
T_dense[4] = T_dense[4] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 3136]);
T_dense[20] = T_dense[20] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 28224]);
T_dense[36] = T_dense[36] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 53312]);
T_dense[52] = T_dense[52] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 78400]);
T_dense[5] = T_dense[5] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 3920]);
T_dense[21] = T_dense[21] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 29008]);
T_dense[37] = T_dense[37] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 54096]);
T_dense[53] = T_dense[53] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 79184]);
T_dense[6] = T_dense[6] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 4704]);
T_dense[22] = T_dense[22] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 29792]);
T_dense[38] = T_dense[38] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 54880]);
T_dense[54] = T_dense[54] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 79968]);
T_dense[7] = T_dense[7] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 5488]);
T_dense[23] = T_dense[23] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 30576]);
T_dense[39] = T_dense[39] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 55664]);
T_dense[55] = T_dense[55] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 80752]);
T_dense[8] = T_dense[8] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 6272]);
T_dense[24] = T_dense[24] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 31360]);
T_dense[40] = T_dense[40] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 56448]);
T_dense[56] = T_dense[56] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 81536]);
T_dense[9] = T_dense[9] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 7056]);
T_dense[25] = T_dense[25] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 32144]);
T_dense[41] = T_dense[41] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 57232]);
T_dense[57] = T_dense[57] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 82320]);
T_dense[10] = T_dense[10] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 7840]);
T_dense[26] = T_dense[26] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 32928]);
T_dense[42] = T_dense[42] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 58016]);
T_dense[58] = T_dense[58] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 83104]);
T_dense[11] = T_dense[11] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 8624]);
T_dense[27] = T_dense[27] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 33712]);
T_dense[43] = T_dense[43] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 58800]);
T_dense[59] = T_dense[59] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 83888]);
T_dense[12] = T_dense[12] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 9408]);
T_dense[28] = T_dense[28] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 34496]);
T_dense[44] = T_dense[44] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 59584]);
T_dense[60] = T_dense[60] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 84672]);
T_dense[13] = T_dense[13] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 10192]);
T_dense[29] = T_dense[29] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 35280]);
T_dense[45] = T_dense[45] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 60368]);
T_dense[61] = T_dense[61] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 85456]);
T_dense[14] = T_dense[14] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 10976]);
T_dense[30] = T_dense[30] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 36064]);
T_dense[46] = T_dense[46] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 61152]);
T_dense[62] = T_dense[62] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 86240]);
T_dense[15] = T_dense[15] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 11760]);
T_dense[31] = T_dense[31] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 36848]);
T_dense[47] = T_dense[47] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 61936]);
T_dense[63] = T_dense[63] + (placeholder_d_shared[index2] * placeholder_shared[index3 + 87024]);
}
}
for (int ax1_inner = 0; ax1_inner < 16; ++ax1_inner) {
int index4 = (int)threadIdx.x * 16 + ax1_inner;
int index5 = ax1_inner;
T_relu[index4] = max(T_dense[index5] + placeholder2[index4], 0.000000e+00f);
T_relu[index4 + 32] = max(T_dense[index5 + 16] + placeholder2[index4 + 32], 0.000000e+00f);
T_relu[index4 + 64] = max(T_dense[index5 + 32] + placeholder2[index4 + 64], 0.000000e+00f);
T_relu[index4 + 96] = max(T_dense[index5 + 48] + placeholder2[index4 + 96], 0.000000e+00f);
}
} |
End of preview. Expand
in Data Studio
Dataset for "Enhancing LLM to Decompile Optimized PTX to Readable CUDA for Tensor Programs" (ASE 2025)
Dataset Description
This dataset accompanies the paper:
Sun X, Tang F, Zhang Y, et al. Enhancing LLM to Decompile Optimized PTX to Readable CUDA for Tensor Programs[C]//2025 40th IEEE/ACM International Conference on Automated Software Engineering (ASE). IEEE, 2025: 2235-2247.
It contains pairs of input PTX kernels and output CUDA kernels generated by two auto-schedulers: Ansor and Welder. The dataset is designed to train and evaluate large language models (LLMs) for the task of decompiling optimized PTX back to human-readable CUDA code.
Dataset Structure
The dataset is split into four Parquet files:
train_ansor.parquettrain_welder.parquettest_ansor.parquettest_welder.parquet
Each file contains samples with the following fields:
| Field | Type | Description |
|---|---|---|
file |
string | Source JSON filename (without extension). For Ansor, it encodes parameters. |
name |
string or int | Sample identifier (string for Ansor, integer for Welder). |
kernel |
string | (Ansor only) Kernel function name. |
input |
string | Input PTX kernel. |
output |
string | Output CUDA kernel. |
Note: The
kernelfield is present only in Ansor samples; Welder samples do not have this field.
Data Splits
| Split | Number of Samples |
|---|---|
| train_ansor | 371,472 |
| train_welder | 383,790 |
| test_ansor | 1,905 |
| test_welder | 1,793 |
Usage Example
from datasets import load_dataset
dataset = load_dataset("your-username/ptxdec-dataset")
train_ansor = dataset["train_ansor"]
print(train_ansor[0])
- Downloads last month
- 7