Texte vergleichen

Finde den Unterschied zwischen zwei Textdateien

Echtzeit-Diff

Einheitlicher Diff

Zeilen einklappen

Änderungen hervorheben

Syntaxhervorhebung

Werkzeuge

Diffchecker Desktop The most secure way to run Diffchecker. Get the Diffchecker Desktop app: your diffs never leave your computer!Get Desktop

matmul-diff-2275

Created 3 years agoDiff never expires

Zeilen
Gesamt
Entfernt

Wörter
Gesamt
Entfernt

Um diese Funktion weiterhin zu nutzen, aktualisieren Sie auf Diffchecker Pro Preise anzeigen

542 Zeilen

Zeilen
Gesamt
Hinzugefügt

Wörter
Gesamt
Hinzugefügt

Um diese Funktion weiterhin zu nutzen, aktualisieren Sie auf Diffchecker Pro Preise anzeigen

480 Zeilen

__global__ void kernel4(Tensor<__half, 2> T0, Tensor<__half, 2> T1, Tensor<float, 2> T4) {

alignas(16) extern __shared__ char array[];

unsigned smem_offset = 0;

NVFUSER_DEFINE_MAGIC_ZERO

int i414;

int i1864;

i414 = ((((((nvfuser_index_t)threadIdx.z) * 2) + ((nvfuser_index_t)threadIdx.y)) * 32) + ((nvfuser_index_t)threadIdx.x)) * 8;

i1864 = (((nvfuser_index_t)threadIdx.z) * 4) + (2 * ((nvfuser_index_t)threadIdx.y));

int i644;

int i1865;

i644 = ((nvfuser_index_t)blockIdx.y) * 128;

i1865 = i1864 + (((nvfuser_index_t)threadIdx.x) / 16);

int i1361;

int i1889;

i1361 = ((nvfuser_index_t)blockIdx.x) * 256;

i1889 = i1865 % 8;

int i3196;

int i1229;

i3196 = ((nvfuser_index_t)threadIdx.x) / 8;

i1229 = ((nvfuser_index_t)threadIdx.x) % 16;

int i1866;

i1866 = i1865 / 8;

int i1875;

i1875 = (i1229 / 8) * 64;

int i1893;

i1893 = i1229 % 8;

int i1896;

i1896 = (((1024 * i1866) + (128 * i1889)) + i1875) + ((i1889 ^ i1893) * 8);

int i3166;

i3166 = T1.size[1] * 8;

int i3192;

i3192 = i3166 * i1866;

int i3193;

i3193 = T1.size[1] * i1889;

int i3180;

i3180 = ((nvfuser_index_t)blockIdx.y) * 128;

int i3197;

i3197 = ((nvfuser_index_t)threadIdx.x) % 8;

i3197 = i1893 * 8;

int i3302;

int i3198;

i3302 = (i3196 / 2) * 8;

i3198 = (((i3192 + i3193) + i3180) + i1875) + i3197;

int i3303;

int i3186;

i3303 = i3302 + i3197;

i3186 = T1.size[1] * 32;

int i3330;

int i3189;

i3330 = i3303 % 8;

i3189 = T1.size[1] * 16;

int i3311;

int i6198;

i3311 = ((nvfuser_index_t)threadIdx.z) * 64;

i6198 = (((nvfuser_index_t)threadIdx.z) * 2) + ((nvfuser_index_t)threadIdx.y);

int i3333;

int i6199;

i3333 = ((((i3303 / 8) * 8) + i3330) * 256) + i3311;

i6199 = i6198 + (((nvfuser_index_t)threadIdx.x) / 32);

int i3318;

int i6221;

i3318 = (i3196 % 2) * 8;

i6221 = i6199 % 8;

int i3520;

int i5800;

i3520 = i3318 + i3197;

i5800 = ((nvfuser_index_t)threadIdx.x) % 32;

int i3523;

int i6200;

i3523 = i3520 % 8;

i6200 = i6199 / 8;

int i3501;

int i6208;

i3501 = ((nvfuser_index_t)threadIdx.y) * 64;

i6208 = (i5800 / 8) * 64;

int i3526;

int i6225;

i3526 = ((((i3520 / 8) * 8) + i3523) * 128) + i3501;

i6225 = i5800 % 8;

int i4604;

int i6228;

i4604 = (i644 + i3501) + ((((nvfuser_index_t)threadIdx.x) % 4) * 2);

i6228 = (((2048 * i6200) + (256 * i6221)) + i6208) + ((i6221 ^ i6225) * 8);

int i4608;

int i7323;

i4608 = (i1361 + i3311) + (((nvfuser_index_t)threadIdx.x) / 4);

i7323 = T0.size[1] * 8;

int i5077;

int i7347;

i5077 = i414 + 7;

i7347 = i7323 * i6200;

int i7348;

i7348 = T0.size[1] * i6221;

int i7335;

i7335 = ((nvfuser_index_t)blockIdx.x) * 256;

int i7352;

i7352 = i6225 * 8;

int i7353;

i7353 = (((i7347 + i7348) + i7335) + i6208) + i7352;

int i7341;

i7341 = T0.size[1] * 32;

int i11742;

i11742 = (((((T1.size[1] * 96) + i3192) + i3193) + i3180) + i1875) + i3197;

int i16008;

i16008 = (((((T0.size[1] * 96) + i7347) + i7348) + i7335) + i6208) + i7352;

int i17482;

i17482 = ((nvfuser_index_t)threadIdx.x) % 8;

int i17724;

i17724 = i17482 % 8;

int i17481;

i17481 = ((nvfuser_index_t)threadIdx.x) / 8;

int i17715;

i17715 = i17482 / 8;

int i17550;

i17550 = i17481 / 2;

int i17829;

i17829 = 2048 * i17550;

int i17830;

i17830 = 2048 * i17715;

int i17832;

i17832 = 256 * i17724;

int i17834;

i17834 = ((nvfuser_index_t)threadIdx.z) * 64;

int i17835;

i17835 = ((i17829 + i17830) + i17832) + i17834;

int i17570;

i17570 = i17481 % 2;

int i18255;

i18255 = 1024 * i17570;

int i18256;

i18256 = 1024 * i17715;

int i18258;

i18258 = 128 * i17724;

int i18246;

i18246 = ((nvfuser_index_t)threadIdx.y) * 64;

int i18260;

i18260 = ((i18255 + i18256) + i18258) + i18246;

int i18918;

i18918 = (((4096 + i17829) + i17830) + i17832) + i17834;

int i19677;

i19677 = (((2048 + i18255) + i18256) + i18258) + i18246;

int i20235;

i20235 = ((nvfuser_index_t)threadIdx.x) / 4;

int i20286;

i20286 = (((nvfuser_index_t)threadIdx.x) % 4) * 2;

int i20301;

i20301 = ((((((T1.size[1] * ((nvfuser_index_t)blockIdx.x)) * 256) + ((T1.size[1] * ((nvfuser_index_t)threadIdx.z)) * 64)) + (T1.size[1] * i20235)) + i3180) + i18246) + i20286;

int i22576;

i22576 = (8 * ((nvfuser_index_t)threadIdx.x)) + 7;

int i22577;

i22577 = i22576 % 128;

int i22578;

i22578 = i22577 / 8;

bool b22625;

b22625 = (((i3180 + ((i22578 / 8) * 64)) + ((i22578 % 8) * 8)) + (i22577 % 8)) < T1.size[1];

int i22627;

i22627 = i1864 + (i22576 / 128);

int i22629;

i22629 = 8 * (i22627 / 8);

int i22630;

i22630 = i22627 % 8;

int i22631;

i22631 = i22629 + i22630;

int i27301;

i27301 = i22576 % 256;

int i27302;

i27302 = i27301 / 8;

bool b27311;

b27311 = (((i7335 + ((i27302 / 8) * 64)) + ((i27302 % 8) * 8)) + (i27301 % 8)) < T0.size[1];

int i27313;

i27313 = i6198 + (i22576 / 256);

int i27315;

i27315 = 8 * (i27313 / 8);

int i27316;

i27316 = i27313 % 8;

int i27317;

i27317 = i27315 + i27316;

int i31979;

i31979 = (96 + i22629) + i22630;

int i36682;

i36682 = (96 + i27315) + i27316;

int i39223;

i39223 = (i7335 + i17834) + i20235;

int i39228;

i39228 = (i3180 + i18246) + i20286;

smem_offset = alignBufferSize(smem_offset, 16);

__half* T7 = reinterpret_cast<__half*>(array + smem_offset);

smem_offset += (((((((ceilDiv(((((ceilDiv(32, 8)) * (ceilDiv(8, 1))) * 1) * (((ceilDiv(128, 64)) * (ceilDiv(64, 8))) * 8)), 2048)) * (ceilDiv((ceilDiv((ceilDiv(2048, 8)), 32)), 2))) * 2) * 32) * 8) * 4) * sizeof(__half));

smem_offset = alignBufferSize(smem_offset, 16);

__half* T6 = reinterpret_cast<__half*>(array + smem_offset);

smem_offset += (((((((ceilDiv(((((ceilDiv(32, 8)) * (ceilDiv(8, 1))) * 1) * (((ceilDiv(256, 64)) * (ceilDiv(64, 8))) * 8)), 2048)) * (ceilDiv((ceilDiv((ceilDiv(2048, 8)), 32)), 2))) * 2) * 32) * 8) * 4) * sizeof(__half));

float T5[(((((ceilDiv(64, 16)) * (ceilDiv(64, 16))) * (ceilDiv(16, 8))) * (ceilDiv(16, 8))) * 2)];

#pragma unroll

for(nvfuser_index_t i166 = 0; i166 < 4; ++i166) {

int i249;

int i251;

i249 = 32 * i166;

i251 = 32 * i166;

#pragma unroll

for(nvfuser_index_t i167 = 0; i167 < 4; ++i167) {

Ampere::initM16N16K16TN<16>(reinterpret_cast<Array<float,8,8>*>(&T5[(i249 + (4 * i167))]));

Ampere::initM16N16K16TN<16>(reinterpret_cast<Array<float,8,8>*>(&T5[(i251 + (4 * i167))]));

}

NVFUSER_UPDATE_MAGIC_ZERO

#pragma unroll

for(nvfuser_index_t i160 = 0; i160 < 3; ++i160) {

int i409;

int i1897;

i409 = 4096 * i160;

i1897 = i1896 + (4096 * i160);

int i679;

int i3199;

i679 = 32 * i160;

i3199 = i3198 + (i3186 * i160);

int i1122;

int i6229;

i1122 = 8192 * i160;

i6229 = i6228 + (8192 * i160);

int i7354;

i7354 = i7353 + (i7341 * i160);

int i22613;

i22613 = 32 * i160;

int i22632;

i22632 = i22631 + i22613;

int i27318;

i27318 = i27317 + i22613;

#pragma unroll

for(nvfuser_index_t i154 = 0; i154 < 2; ++i154) {

int i416;

bool b22635;

i416 = i414 + (2048 * i154);

b22635 = b22625 && ((i22632 + (16 * (i154 + nvfuser_zero))) < T0.size[0]);

int i417;

if ((!b22635)) {

i417 = i416 / 128;

arraySet<__half, 8>(&T7[(i1897 + (2048 * i154))], (__half)0);

int i476;

i476 = i417 % 8;

int i480;

i480 = i416 % 128;

int i481;

i481 = i480 / 8;

int i5078;

i5078 = i5077 + (2048 * (i154 + nvfuser_zero));

int i5079;

i5079 = i5078 % 128;

int i5080;

i5080 = i5079 / 8;

int i5090;

i5090 = i5078 / 128;

bool b5097;

b5097 = ((((i644 + (64 * (i5080 / 8))) + (8 * (i5080 % 8))) + (i5079 % 8)) < T1.size[1]) && (((i679 + (8 * (i5090 / 8))) + (i5090 % 8)) < T0.size[0]);

if ((!b5097)) {

arraySet<__half, 8>(&T7[((((i409 + (128 * ((8 * (i417 / 8)) + i476))) + (64 * (i481 / 8))) + (8 * (i476 ^ (i481 % 8)))) + (i480 % 8))], (__half)0);

}

#pragma unroll

for(nvfuser_index_t i154 = 0; i154 < 2; ++i154) {

int i650;

int i1912;

i650 = 2048 * (i154 + nvfuser_zero);

i1912 = i154 + nvfuser_zero;

int i703;

Ampere::cpAsync(reinterpret_cast<Array<__half,8,8>*>(&T7[(i1897 + (2048 * i154))]),reinterpret_cast<Array<__half,8,8>*>(&T1[(i3199 + (i3189 * i1912))]),(b22625 && ((i22632 + (16 * i1912)) < T0.size[0])));

i703 = i414 + i650;

int i704;

i704 = i703 % 128;

int i705;

i705 = i704 / 8;

int i714;

i714 = i703 / 128;

int i946;

i946 = i414 + (2048 * i154);

int i947;

i947 = i946 / 128;

int i950;

i950 = i947 % 8;

int i954;

i954 = i946 % 128;

int i955;

i955 = i954 / 8;

int i5695;

i5695 = i5077 + i650;

int i5696;

i5696 = i5695 % 128;

int i5697;

i5697 = i5696 / 8;

int i5707;

i5707 = i5695 / 128;

Ampere::cpAsync(reinterpret_cast<Array<__half,8,8>*>(&T7[((((i409 + (128 * ((8 * (i947 / 8)) + i950))) + (64 * (i955 / 8))) + (8 * (i950 ^ (i955 % 8)))) + (i954 % 8))]),reinterpret_cast<Array<__half,8,8>*>(&T1[((((i644 + (64 * (i705 / 8))) + (8 * (i705 % 8))) + (i704 % 8)) + (T1.size[1] * ((i679 + (8 * (i714 / 8))) + (i714 % 8))))]),(((((i644 + (64 * (i5697 / 8))) + (8 * (i5697 % 8))) + (i5696 % 8)) < T1.size[1]) && (((i679 + (8 * (i5707 / 8))) + (i5707 % 8)) < T0.size[0])));

}

#pragma unroll

for(nvfuser_index_t i143 = 0; i143 < 4; ++i143) {

int i1189;

bool b27321;

i1189 = i414 + (2048 * i143);

b27321 = b27311 && ((i27318 + (8 * (i143 + nvfuser_zero))) < T0.size[0]);

int i1190;

if ((!b27321)) {

i1190 = i1189 / 256;

arraySet<__half, 8>(&T6[(i6229 + (2048 * i143))], (__half)0);

int i1193;

i1193 = i1190 % 8;

int i1197;

i1197 = i1189 % 256;

int i1198;

i1198 = i1197 / 8;

int i6178;

i6178 = i5077 + (2048 * (i143 + nvfuser_zero));

int i6179;

i6179 = i6178 % 256;

int i6180;

i6180 = i6179 / 8;

int i6190;

i6190 = i6178 / 256;

bool b6197;

b6197 = ((((i1361 + (64 * (i6180 / 8))) + (8 * (i6180 % 8))) + (i6179 % 8)) < T0.size[1]) && (((i679 + (8 * (i6190 / 8))) + (i6190 % 8)) < T0.size[0]);

if ((!b6197)) {

arraySet<__half, 8>(&T6[((((i1122 + (256 * ((8 * (i1190 / 8)) + i1193))) + (64 * (i1198 / 8))) + (8 * (i1193 ^ (i1198 % 8)))) + (i1197 % 8))], (__half)0);

}

#pragma unroll

for(nvfuser_index_t i143 = 0; i143 < 4; ++i143) {

int i1367;

int i6244;

i1367 = 2048 * (i143 + nvfuser_zero);

i6244 = i143 + nvfuser_zero;

int i1420;

Ampere::cpAsync(reinterpret_cast<Array<__half,8,8>*>(&T6[(i6229 + (2048 * i143))]),reinterpret_cast<Array<__half,8,8>*>(&T0[(i7354 + (i7323 * i6244))]),(b27311 && ((i27318 + (8 * i6244)) < T0.size[0])));

i1420 = i414 + i1367;

int i1421;

i1421 = i1420 % 256;

int i1422;

i1422 = i1421 / 8;

int i1431;

i1431 = i1420 / 256;

int i1663;

i1663 = i414 + (2048 * i143);

int i1664;

i1664 = i1663 / 256;

int i1667;

i1667 = i1664 % 8;

int i1671;

i1671 = i1663 % 256;

int i1672;

i1672 = i1671 / 8;

int i6795;

i6795 = i5077 + i1367;

int i6796;

i6796 = i6795 % 256;

int i6797;

i6797 = i6796 / 8;

int i6807;

i6807 = i6795 / 256;

Ampere::cpAsync(reinterpret_cast<Array<__half,8,8>*>(&T6[((((i1122 + (256 * ((8 * (i1664 / 8)) + i1667))) + (64 * (i1672 / 8))) + (8 * (i1667 ^ (i1672 % 8)))) + (i1671 % 8))]),reinterpret_cast<Array<__half,8,8>*>(&T0[((((i1361 + (64 * (i1422 / 8))) + (8 * (i1422 % 8))) + (i1421 % 8)) + (T0.size[1] * ((i679 + (8 * (i1431 / 8))) + (i1431 % 8))))]),(((((i1361 + (64 * (i6797 / 8))) + (8 * (i6797 % 8))) + (i6796 % 8)) < T0.size[1]) && (((i679 + (8 * (i6807 / 8))) + (i6807 % 8)) < T0.size[0])));

}

Ampere::cpAsyncCommit();

}

NVFUSER_UPDATE_MAGIC_ZERO

Ampere::cpAsyncPartialBarrier<2>();

__barrier_sync(0);

#pragma unroll 1

for(nvfuser_index_t i161 = 0; i161 < (ceilDiv(T0.size[0], 32)); ++i161) {

int i1846;

int i10426;

i1846 = 3 + i161;

i10426 = (3 + i161) % 4;

int i1847;

int i10431;

i1847 = i1846 % 4;

i10431 = i1896 + (4096 * i10426);

int i1848;

int i11744;

i1848 = 4096 * i1847;

i11744 = i11742 + (i3186 * i161);

int i2162;

int i14873;

i2162 = 32 * i1846;

i14873 = i6228 + (8192 * i10426);

int i2656;

int i16010;

i2656 = 8192 * i1847;

i16010 = i16008 + (i7341 * i161);

int i3259;

int i17544;

i3259 = i161 % 4;

i17544 = i161 % 4;

int i3313;

int i17836;

i3313 = 8192 * i3259;

i17836 = 8192 * i17544;

int i3334;

int i17837;

i3334 = i3333 + i3313;

i17837 = i17835 + i17836;

int i3527;

int i18261;

i3527 = 4096 * i3259;

i18261 = 4096 * i17544;

int i3528;

int i18262;

i3528 = i3526 + i3527;

i18262 = i18260 + i18261;

int i3743;

int i18919;

i3743 = i3311 + i3313;

i18919 = i18918 + i17836;

int i4054;

int i19678;

i4054 = i3501 + i3527;

i19678 = i19677 + i18261;

int i31972;

i31972 = 32 * i161;

int i31980;

i31980 = i31979 + i31972;

int i36683;

i36683 = i36682 + i31972;

#pragma unroll

for(nvfuser_index_t i154 = 0; i154 < 2; ++i154) {

int i1915;

bool b31983;

i1915 = i414 + (2048 * i154);

b31983 = b22625 && ((i31980 + (16 * (i154 + nvfuser_zero))) < T0.size[0]);

int i1916;

if ((!b31983)) {

i1916 = i1915 / 128;

arraySet<__half, 8>(&T7[(i10431 + (2048 * i154))], (__half)0);

int i1919;

i1919 = i1916 % 8;

int i1923;

i1923 = i1915 % 128;

int i1924;

i1924 = i1923 / 8;

int i7288;

i7288 = i5077 + (2048 * (i154 + nvfuser_zero));

int i7289;

i7289 = i7288 % 128;

int i7290;

i7290 = i7289 / 8;

int i7300;

i7300 = i7288 / 128;

bool b7307;

b7307 = ((((i644 + (64 * (i7290 / 8))) + (8 * (i7290 % 8))) + (i7289 % 8)) < T1.size[1]) && (((i2162 + (8 * (i7300 / 8))) + (i7300 % 8)) < T0.size[0]);

if ((!b7307)) {

arraySet<__half, 8>(&T7[((((i1848 + (128 * ((8 * (i1916 / 8)) + i1919))) + (64 * (i1924 / 8))) + (8 * (i1919 ^ (i1924 % 8)))) + (i1923 % 8))], (__half)0);

}

NVFUSER_UPDATE_MAGIC_ZERO

#pragma unroll

for(nvfuser_index_t i154 = 0; i154 < 2; ++i154) {

int i2097;

int i10446;

i2097 = 2048 * (i154 + nvfuser_zero);

i10446 = i154 + nvfuser_zero;

int i2151;

Ampere::cpAsync(reinterpret_cast<Array<__half,8,8>*>(&T7[(i10431 + (2048 * i154))]),reinterpret_cast<Array<__half,8,8>*>(&T1[(i11744 + (i3189 * i10446))]),(b22625 && ((i31980 + (16 * i10446)) < T0.size[0])));

i2151 = i414 + i2097;

int i2152;

i2152 = i2151 % 128;

int i2153;

i2153 = i2152 / 8;

int i2163;

i2163 = i2151 / 128;

int i2404;

i2404 = i414 + (2048 * i154);

int i2405;

i2405 = i2404 / 128;

int i2408;

i2408 = i2405 % 8;

int i2412;

i2412 = i2404 % 128;

int i2413;

i2413 = i2412 / 8;

int i7917;

i7917 = i5077 + i2097;

int i7918;

i7918 = i7917 % 128;

int i7919;

i7919 = i7918 / 8;

int i7929;

i7929 = i7917 / 128;

Ampere::cpAsync(reinterpret_cast<Array<__half,8,8>*>(&T7[((((i1848 + (128 * ((8 * (i2405 / 8)) + i2408))) + (64 * (i2413 / 8))) + (8 * (i2408 ^ (i2413 % 8)))) + (i2412 % 8))]),reinterpret_cast<Array<__half,8,8>*>(&T1[((((i644 + (64 * (i2153 / 8))) + (8 * (i2153 % 8))) + (i2152 % 8)) + (T1.size[1] * ((i2162 + (8 * (i2163 / 8))) + (i2163 % 8))))]),(((((i644 + (64 * (i7919 / 8))) + (8 * (i7919 % 8))) + (i7918 % 8)) < T1.size[1]) && (((i2162 + (8 * (i7929 / 8))) + (i7929 % 8)) < T0.size[0])));

}

NVFUSER_UPDATE_MAGIC_ZERO

#pragma unroll

for(nvfuser_index_t i143 = 0; i143 < 4; ++i143) {

int i2657;

bool b36686;

i2657 = i414 + (2048 * i143);

b36686 = b27311 && ((i36683 + (8 * (i143 + nvfuser_zero))) < T0.size[0]);

int i2658;

if ((!b36686)) {

i2658 = i2657 / 256;

arraySet<__half, 8>(&T6[(i14873 + (2048 * i143))], (__half)0);

int i2661;

i2661 = i2658 % 8;

int i2665;

i2665 = i2657 % 256;

int i2666;

i2666 = i2665 / 8;

int i8410;

i8410 = i5077 + (2048 * (i143 + nvfuser_zero));

int i8411;

i8411 = i8410 % 256;

int i8412;

i8412 = i8411 / 8;

int i8422;

i8422 = i8410 / 256;

bool b8429;

b8429 = ((((i1361 + (64 * (i8412 / 8))) + (8 * (i8412 % 8))) + (i8411 % 8)) < T0.size[1]) && (((i2162 + (8 * (i8422 / 8))) + (i8422 % 8)) < T0.size[0]);

if ((!b8429)) {

arraySet<__half, 8>(&T6[((((i2656 + (256 * ((8 * (i2658 / 8)) + i2661))) + (64 * (i2666 / 8))) + (8 * (i2661 ^ (i2666 % 8)))) + (i2665 % 8))], (__half)0);

}

NVFUSER_UPDATE_MAGIC_ZERO

#pragma unroll

for(nvfuser_index_t i143 = 0; i143 < 4; ++i143) {

int i2839;

int i14888;

i2839 = 2048 * (i143 + nvfuser_zero);

i14888 = i143 + nvfuser_zero;

int i2893;

Ampere::cpAsync(reinterpret_cast<Array<__half,8,8>*>(&T6[(i14873 + (2048 * i143))]),reinterpret_cast<Array<__half,8,8>*>(&T0[(i16010 + (i7323 * i14888))]),(b27311 && ((i36683 + (8 * i14888)) < T0.size[0])));

i2893 = i414 + i2839;

int i2894;

i2894 = i2893 % 256;

int i2895;

i2895 = i2894 / 8;

int i2904;

i2904 = i2893 / 256;

int i3145;

i3145 = i414 + (2048 * i143);

int i3146;

i3146 = i3145 / 256;

int i3149;

i3149 = i3146 % 8;

int i3153;

i3153 = i3145 % 256;

int i3154;

i3154 = i3153 / 8;

int i9039;

i9039 = i5077 + i2839;

int i9040;

i9040 = i9039 % 256;

int i9041;

i9041 = i9040 / 8;

int i9051;

i9051 = i9039 / 256;

Ampere::cpAsync(reinterpret_cast<Array<__half,8,8>*>(&T6[((((i2656 + (256 * ((8 * (i3146 / 8)) + i3149))) + (64 * (i3154 / 8))) + (8 * (i3149 ^ (i3154 % 8)))) + (i3153 % 8))]),reinterpret_cast<Array<__half,8,8>*>(&T0[((((i1361 + (64 * (i2895 / 8))) + (8 * (i2895 % 8))) + (i2894 % 8)) + (T0.size[1] * ((i2162 + (8 * (i2904 / 8))) + (i2904 % 8))))]),(((((i1361 + (64 * (i9041 / 8))) + (8 * (i9041 % 8))) + (i9040 % 8)) < T0.size[1]) && (((i2162 + (8 * (i9051 / 8))) + (i9051 % 8)) < T0.size[0])));

}

NVFUSER_UPDATE_MAGIC_ZERO

Array<__half, (((ceilDiv(64, 16)) * 8) * 2), 8> T8;

Array<__half, (((ceilDiv(64, 16)) * 8) * 2), 8> T9;

#pragma unroll

for(nvfuser_index_t i145 = 0; i145 < 4; ++i145) {

int i3320;

Turing::ldMatrixT (*reinterpret_cast<Array<__half,8,8>*>(&T8[(8 * i145)]),&T6[(i17837 + (8 * (i17724 ^ (i17570 + (2 * i145)))))]);

i3320 = i3318 + (16 * i145);

Turing::ldMatrixT (*reinterpret_cast<Array<__half,8,8>*>(&T8[(8 * i145)]),&T6[((i3334 + (8 * (i3330 ^ (i3320 / 8)))) + (i3320 % 8))]);

}

NVFUSER_UPDATE_MAGIC_ZERO

#pragma unroll

for(nvfuser_index_t i156 = 0; i156 < 4; ++i156) {

int i3529;

Turing::ldMatrixT (*reinterpret_cast<Array<__half,8,8>*>(&T9[(8 * i156)]),&T7[(i18262 + (8 * (i17724 ^ (i17550 + (2 * i156)))))]);

i3529 = i3302 + (16 * i156);

Turing::ldMatrixT (*reinterpret_cast<Array<__half,8,8>*>(&T9[(8 * i156)]),&T7[((i3528 + (8 * (i3523 ^ (i3529 / 8)))) + (i3529 % 8))]);

}

NVFUSER_UPDATE_MAGIC_ZERO

#pragma unroll

for(nvfuser_index_t i164 = 0; i164 < 1; ++i164) {

int i3711;

int i18920;

i3711 = 1 + i164;

i18920 = i18919 + (4096 * i164);

int i3712;

int i18960;

i3712 = 16 * i3711;

i18960 = 32 * ((1 + i164) % 2);

int i3744;

int i19035;

i3744 = i3303 + i3712;

i19035 = 32 * (i164 % 2);

int i3747;

int i19679;

i3747 = i3744 % 8;

i19679 = i19678 + (2048 * i164);

int i3750;

i3750 = i3743 + (256 * ((8 * (i3744 / 8)) + i3747));

int i3789;

i3789 = 32 * (i3711 % 2);

int i3855;

i3855 = 32 * (i164 % 2);

int i4055;

i4055 = i3520 + i3712;

int i4058;

i4058 = i4055 % 8;

int i4061;

i4061 = i4054 + (128 * ((8 * (i4055 / 8)) + i4058));

#pragma unroll

for(nvfuser_index_t i145 = 0; i145 < 4; ++i145) {

int i3751;

Turing::ldMatrixT (*reinterpret_cast<Array<__half,8,8>*>(&T8[(i18960 + (8 * i145))]),&T6[(i18920 + (8 * (i17724 ^ (i17570 + (2 * i145)))))]);

i3751 = i3318 + (16 * i145);

Turing::ldMatrixT (*reinterpret_cast<Array<__half,8,8>*>(&T8[(i3789 + (8 * i145))]),&T6[((i3750 + (i3751 % 8)) + (8 * (i3747 ^ (i3751 / 8))))]);

}

__half T2[((ceilDiv(64, 16)) * 8)];

#pragma unroll

for(nvfuser_index_t i147 = 0; i147 < 4; ++i147) {

int i3808;

int i18982;

i3808 = 8 * i147;

i18982 = 8 * i147;

#pragma unroll

for(nvfuser_index_t i149 = 0; i149 < 8; ++i149) {

T2[(i3808 + i149)] = 0.00000000000000000e+00;

T2[(i18982 + i149)] = 0.00000000000000000e+00;

}

#pragma unroll

for(nvfuser_index_t i147 = 0; i147 < 4; ++i147) {

int i3827;

int i19003;

i3827 = 8 * i147;

i19003 = 8 * i147;

int i3859;

int i19039;

i3859 = i3855 + i3827;

i19039 = i19035 + i19003;

#pragma unroll

for(nvfuser_index_t i149 = 0; i149 < 8; ++i149) {

T2[(i3827 + i149)]

T2[(i19003 + i149)]

= T8[(i3859 + i149)];

= T8[(i19039 + i149)];

}

#pragma unroll

for(nvfuser_index_t i156 = 0; i156 < 4; ++i156) {

int i4062;

Turing::ldMatrixT (*reinterpret_cast<Array<__half,8,8>*>(&T9[(i18960 + (8 * i156))]),&T7[(i19679 + (8 * (i17724 ^ (i17550 + (2 * i156)))))]);

i4062 = i3302 + (16 * i156);

Turing::ldMatrixT (*reinterpret_cast<Array<__half,8,8>*>(&T9[(i3789 + (8 * i156))]),&T7[((i4061 + (i4062 % 8)) + (8 * (i4058 ^ (i4062 / 8))))]);

}

__half T3[((ceilDiv(64, 16)) * 8)];

#pragma unroll

for(nvfuser_index_t i158 = 0; i158 < 4; ++i158) {

int i4117;

int i19742;

i4117 = 8 * i158;

i19742 = 8 * i158;

#pragma unroll

for(nvfuser_index_t i159 = 0; i159 < 8; ++i159) {

T3[(i4117 + i159)] = 0.00000000000000000e+00;

T3[(i19742 + i159)] = 0.00000000000000000e+00;

}

#pragma unroll

for(nvfuser_index_t i158 = 0; i158 < 4; ++i158) {

int i4136;

int i19763;

i4136 = 8 * i158;

i19763 = 8 * i158;

int i4168;

int i19799;

i4168 = i3855 + i4136;

i19799 = i19035 + i19763;

#pragma unroll

for(nvfuser_index_t i159 = 0; i159 < 8; ++i159) {

T3[(i4136 + i159)]

T3[(i19763 + i159)]

= T9[(i4168 + i159)];

= T9[(i19799 + i159)];

}

#pragma unroll

for(nvfuser_index_t i166 = 0; i166 < 4; ++i166) {

int i4226;

int i19866;

i4226 = 32 * i166;

i19866 = 32 * i166;

#pragma unroll

for(nvfuser_index_t i167 = 0; i167 < 4; ++i167) {

Ampere::M16N16K16TN<16>(

reinterpret_cast<Array<float,8,8>*>(&T5[(i4226 + (4 * i167))]),

reinterpret_cast<Array<float,8,8>*>(&T5[(i19866 + (4 * i167))]),

&(reinterpret_cast<Array<__half,8,8>*>(&T2)[i166]),

&(reinterpret_cast<Array<__half,8,8>*>(&T3)[i167]));

}

NVFUSER_UPDATE_MAGIC_ZERO

__half T2[((ceilDiv(64, 16)) * 8)];

#pragma unroll

for(nvfuser_index_t i147 = 0; i147 < 4; ++i147) {

int i4246;

int i19887;

i4246 = 8 * i147;

i19887 = 8 * i147;

#pragma unroll

for(nvfuser_index_t i149 = 0; i149 < 8; ++i149) {

T2[(i4246 + i149)] = 0.00000000000000000e+00;

T2[(i19887 + i149)] = 0.00000000000000000e+00;

}

#pragma unroll

for(nvfuser_index_t i147 = 0; i147 < 4; ++i147) {

int i4265;

int i19907;

i4265 = 8 * i147;

i19907 = 8 * i147;

int i4295;

int i19938;

i4295 = 32 + i4265;

i19938 = 32 + i19907;

#pragma unroll

for(nvfuser_index_t i149 = 0; i149 < 8; ++i149) {

T2[(i4265 + i149)]

T2[(i19907 + i149)]

= T8[(i4295 + i149)];

= T8[(i19938 + i149)];

}

__half T3[((ceilDiv(64, 16)) * 8)];

#pragma unroll

for(nvfuser_index_t i158 = 0; i158 < 4; ++i158) {

int i4314;

int i19958;

i4314 = 8 * i158;

i19958 = 8 * i158;

#pragma unroll

for(nvfuser_index_t i159 = 0; i159 < 8; ++i159) {

T3[(i4314 + i159)] = 0.00000000000000000e+00;

T3[(i19958 + i159)] = 0.00000000000000000e+00;

}

#pragma unroll

for(nvfuser_index_t i158 = 0; i158 < 4; ++i158) {

int i4333;

int i19978;

i4333 = 8 * i158;

i19978 = 8 * i158;

int i4363;

int i20009;

i4363 = 32 + i4333;

i20009 = 32 + i19978;

#pragma unroll

for(nvfuser_index_t i159 = 0; i159 < 8; ++i159) {

T3[(i4333 + i159)]

T3[(i19978 + i159)]

= T9[(i4363 + i159)];

= T9[(i20009 + i159)];

}

#pragma unroll

for(nvfuser_index_t i166 = 0; i166 < 4; ++i166) {

int i4421;

int i20073;

i4421 = 32 * i166;

i20073 = 32 * i166;

#pragma unroll

for(nvfuser_index_t i167 = 0; i167 < 4; ++i167) {

Ampere::M16N16K16TN<16>(

reinterpret_cast<Array<float,8,8>*>(&T5[(i4421 + (4 * i167))]),

reinterpret_cast<Array<float,8,8>*>(&T5[(i20073 + (4 * i167))]),

&(reinterpret_cast<Array<__half,8,8>*>(&T2)[i166]),

&(reinterpret_cast<Array<__half,8,8>*>(&T3)[i167]));

}

NVFUSER_UPDATE_MAGIC_ZERO

Ampere::cpAsyncPartialBarrier<2>();

__barrier_sync(0);

Ampere::cpAsyncCommit();

}

#pragma unroll

for(nvfuser_index_t i172 = 0; i172 < 4; ++i172) {

int i4475;

int i20139;

i4475 = 2 * i172;

i20139 = 32 * i172;

int i4609;

int i20303;

i4609 = i4608 + (16 * i172);

i20303 = i20301 + (i3189 * i172);

int i39224;

i39224 = i39223 + (16 * i172);

#pragma unroll

for(nvfuser_index_t i173 = 0; i173 < 4; ++i173) {

int i4472;

int i20141;

i4472 = 4 * i173;

i20141 = i20139 + (4 * i173);

int i4605;

int i20291;

i4605 = i4604 + (16 * i173);

i20291 = 16 * i173;

int i20304;

i20304 = i20303 + i20291;

int i39229;

i39229 = i39228 + i20291;

#pragma unroll

for(nvfuser_index_t i174 = 0; i174 < 2; ++i174) {

int i4474;

int i20143;

i4474 = i4472 + (2 * i174);

i20143 = i20141 + (2 * i174);

int i4606;

int i20293;

i4606 = i4605 + (8 * i174);

i20293 = 8 * i174;

int i20305;

i20305 = i20304 + i20293;

int i39230;

i39230 = i39229 + i20293;

#pragma unroll

for(nvfuser_index_t i175 = 0; i175 < 2; ++i175) {

int i4478;

int i20145;

i4478 = i4474 + (16 * (i4475 + i175));

i20145 = i20143 + (16 * i175);

int i4612;

int i20307;

i4612 = i4606 + (T1.size[1] * (i4609 + (8 * i175)));

i20307 = i20305 + (i3166 * i175);

bool b9214;

bool b39226;

b9214 = (i4609 + (8 * (i175 + nvfuser_zero))) < T0.size[1];

b39226 = (i39224 + (8 * (i175 + nvfuser_zero))) < T0.size[1];

#pragma unroll

for(nvfuser_index_t i176 = 0; i176 < 2; ++i176) {

int i4480;

int i20147;

i4480 = i176 + nvfuser_zero;

i20147 = i176 + nvfuser_zero;

if ((b9214 && ((i4606 + i4480) < T1.size[1]))) {

if ((b39226 && ((i39230 + i20147) < T1.size[1]))) {

T4[(i4612 + i4480)]

T4[(i20307 + i20147)]

= T5[(i4478 + i176)];

= T5[(i20145 + i176)];

}

NVFUSER_UPDATE_MAGIC_ZERO

}

Gespeicherte Diffs

Originaltext

Datei öffnen

__global__ void kernel4(Tensor<__half, 2> T0, Tensor<__half, 2> T1, Tensor<float, 2> T4) {
  alignas(16) extern __shared__ char array[];
  unsigned smem_offset = 0;
  NVFUSER_DEFINE_MAGIC_ZERO
  int i414;
  i414 = ((((((nvfuser_index_t)threadIdx.z) * 2) + ((nvfuser_index_t)threadIdx.y)) * 32) + ((nvfuser_index_t)threadIdx.x)) * 8;
  int i644;
  i644 = ((nvfuser_index_t)blockIdx.y) * 128;
  int i1361;
  i1361 = ((nvfuser_index_t)blockIdx.x) * 256;
  int i3196;
  i3196 = ((nvfuser_index_t)threadIdx.x) / 8;
  int i3197;
  i3197 = ((nvfuser_index_t)threadIdx.x) % 8;
  int i3302;
  i3302 = (i3196 / 2) * 8;
  int i3303;
  i3303 = i3302 + i3197;
  int i3330;
  i3330 = i3303 % 8;
  int i3311;
  i3311 = ((nvfuser_index_t)threadIdx.z) * 64;
  int i3333;
  i3333 = ((((i3303 / 8) * 8) + i3330) * 256) + i3311;
  int i3318;
  i3318 = (i3196 % 2) * 8;
  int i3520;
  i3520 = i3318 + i3197;
  int i3523;
  i3523 = i3520 % 8;
  int i3501;
  i3501 = ((nvfuser_index_t)threadIdx.y) * 64;
  int i3526;
  i3526 = ((((i3520 / 8) * 8) + i3523) * 128) + i3501;
  int i4604;
  i4604 = (i644 + i3501) + ((((nvfuser_index_t)threadIdx.x) % 4) * 2);
  int i4608;
  i4608 = (i1361 + i3311) + (((nvfuser_index_t)threadIdx.x) / 4);
  int i5077;
  i5077 = i414 + 7;
  smem_offset = alignBufferSize(smem_offset, 16);
  __half* T7 = reinterpret_cast<__half*>(array + smem_offset);
  smem_offset += (((((((ceilDiv(((((ceilDiv(32, 8)) * (ceilDiv(8, 1))) * 1) * (((ceilDiv(128, 64)) * (ceilDiv(64, 8))) * 8)), 2048)) * (ceilDiv((ceilDiv((ceilDiv(2048, 8)), 32)), 2))) * 2) * 32) * 8) * 4) * sizeof(__half));
  smem_offset = alignBufferSize(smem_offset, 16);
  __half* T6 = reinterpret_cast<__half*>(array + smem_offset);
  smem_offset += (((((((ceilDiv(((((ceilDiv(32, 8)) * (ceilDiv(8, 1))) * 1) * (((ceilDiv(256, 64)) * (ceilDiv(64, 8))) * 8)), 2048)) * (ceilDiv((ceilDiv((ceilDiv(2048, 8)), 32)), 2))) * 2) * 32) * 8) * 4) * sizeof(__half));
  float T5[(((((ceilDiv(64, 16)) * (ceilDiv(64, 16))) * (ceilDiv(16, 8))) * (ceilDiv(16, 8))) * 2)];
  #pragma unroll
  for(nvfuser_index_t i166 = 0; i166 < 4; ++i166) {
    int i249;
    i249 = 32 * i166;
    #pragma unroll
    for(nvfuser_index_t i167 = 0; i167 < 4; ++i167) {
      Ampere::initM16N16K16TN<16>(reinterpret_cast<Array<float,8,8>*>(&T5[(i249 + (4 * i167))]));
    }
  }
  NVFUSER_UPDATE_MAGIC_ZERO
  #pragma unroll
  for(nvfuser_index_t i160 = 0; i160 < 3; ++i160) {
    int i409;
    i409 = 4096 * i160;
    int i679;
    i679 = 32 * i160;
    int i1122;
    i1122 = 8192 * i160;
    #pragma unroll
    for(nvfuser_index_t i154 = 0; i154 < 2; ++i154) {
      int i416;
      i416 = i414 + (2048 * i154);
      int i417;
      i417 = i416 / 128;
      int i476;
      i476 = i417 % 8;
      int i480;
      i480 = i416 % 128;
      int i481;
      i481 = i480 / 8;
      int i5078;
      i5078 = i5077 + (2048 * (i154 + nvfuser_zero));
      int i5079;
      i5079 = i5078 % 128;
      int i5080;
      i5080 = i5079 / 8;
      int i5090;
      i5090 = i5078 / 128;
      bool b5097;
      b5097 = ((((i644 + (64 * (i5080 / 8))) + (8 * (i5080 % 8))) + (i5079 % 8)) < T1.size[1]) && (((i679 + (8 * (i5090 / 8))) + (i5090 % 8)) < T0.size[0]);
      if ((!b5097)) {
        arraySet<__half, 8>(&T7[((((i409 + (128 * ((8 * (i417 / 8)) + i476))) + (64 * (i481 / 8))) + (8 * (i476 ^ (i481 % 8)))) + (i480 % 8))], (__half)0);
      }
    }
    #pragma unroll
    for(nvfuser_index_t i154 = 0; i154 < 2; ++i154) {
      int i650;
      i650 = 2048 * (i154 + nvfuser_zero);
      int i703;
      i703 = i414 + i650;
      int i704;
      i704 = i703 % 128;
      int i705;
      i705 = i704 / 8;
      int i714;
      i714 = i703 / 128;
      int i946;
      i946 = i414 + (2048 * i154);
      int i947;
      i947 = i946 / 128;
      int i950;
      i950 = i947 % 8;
      int i954;
      i954 = i946 % 128;
      int i955;
      i955 = i954 / 8;
      int i5695;
      i5695 = i5077 + i650;
      int i5696;
      i5696 = i5695 % 128;
      int i5697;
      i5697 = i5696 / 8;
      int i5707;
      i5707 = i5695 / 128;
      Ampere::cpAsync(reinterpret_cast<Array<__half,8,8>*>(&T7[((((i409 + (128 * ((8 * (i947 / 8)) + i950))) + (64 * (i955 / 8))) + (8 * (i950 ^ (i955 % 8)))) + (i954 % 8))]),reinterpret_cast<Array<__half,8,8>*>(&T1[((((i644 + (64 * (i705 / 8))) + (8 * (i705 % 8))) + (i704 % 8)) + (T1.size[1] * ((i679 + (8 * (i714 / 8))) + (i714 % 8))))]),(((((i644 + (64 * (i5697 / 8))) + (8 * (i5697 % 8))) + (i5696 % 8)) < T1.size[1]) && (((i679 + (8 * (i5707 / 8))) + (i5707 % 8)) < T0.size[0])));
    }
    #pragma unroll
    for(nvfuser_index_t i143 = 0; i143 < 4; ++i143) {
      int i1189;
      i1189 = i414 + (2048 * i143);
      int i1190;
      i1190 = i1189 / 256;
      int i1193;
      i1193 = i1190 % 8;
      int i1197;
      i1197 = i1189 % 256;
      int i1198;
      i1198 = i1197 / 8;
      int i6178;
      i6178 = i5077 + (2048 * (i143 + nvfuser_zero));
      int i6179;
      i6179 = i6178 % 256;
      int i6180;
      i6180 = i6179 / 8;
      int i6190;
      i6190 = i6178 / 256;
      bool b6197;
      b6197 = ((((i1361 + (64 * (i6180 / 8))) + (8 * (i6180 % 8))) + (i6179 % 8)) < T0.size[1]) && (((i679 + (8 * (i6190 / 8))) + (i6190 % 8)) < T0.size[0]);
      if ((!b6197)) {
        arraySet<__half, 8>(&T6[((((i1122 + (256 * ((8 * (i1190 / 8)) + i1193))) + (64 * (i1198 / 8))) + (8 * (i1193 ^ (i1198 % 8)))) + (i1197 % 8))], (__half)0);
      }
    }
    #pragma unroll
    for(nvfuser_index_t i143 = 0; i143 < 4; ++i143) {
      int i1367;
      i1367 = 2048 * (i143 + nvfuser_zero);
      int i1420;
      i1420 = i414 + i1367;
      int i1421;
      i1421 = i1420 % 256;
      int i1422;
      i1422 = i1421 / 8;
      int i1431;
      i1431 = i1420 / 256;
      int i1663;
      i1663 = i414 + (2048 * i143);
      int i1664;
      i1664 = i1663 / 256;
      int i1667;
      i1667 = i1664 % 8;
      int i1671;
      i1671 = i1663 % 256;
      int i1672;
      i1672 = i1671 / 8;
      int i6795;
      i6795 = i5077 + i1367;
      int i6796;
      i6796 = i6795 % 256;
      int i6797;
      i6797 = i6796 / 8;
      int i6807;
      i6807 = i6795 / 256;
      Ampere::cpAsync(reinterpret_cast<Array<__half,8,8>*>(&T6[((((i1122 + (256 * ((8 * (i1664 / 8)) + i1667))) + (64 * (i1672 / 8))) + (8 * (i1667 ^ (i1672 % 8)))) + (i1671 % 8))]),reinterpret_cast<Array<__half,8,8>*>(&T0[((((i1361 + (64 * (i1422 / 8))) + (8 * (i1422 % 8))) + (i1421 % 8)) + (T0.size[1] * ((i679 + (8 * (i1431 / 8))) + (i1431 % 8))))]),(((((i1361 + (64 * (i6797 / 8))) + (8 * (i6797 % 8))) + (i6796 % 8)) < T0.size[1]) && (((i679 + (8 * (i6807 / 8))) + (i6807 % 8)) < T0.size[0])));
    }
    Ampere::cpAsyncCommit();
  }
  NVFUSER_UPDATE_MAGIC_ZERO
  Ampere::cpAsyncPartialBarrier<2>();
  __barrier_sync(0);
  #pragma unroll 1
  for(nvfuser_index_t i161 = 0; i161 < (ceilDiv(T0.size[0], 32)); ++i161) {
    int i1846;
    i1846 = 3 + i161;
    int i1847;
    i1847 = i1846 % 4;
    int i1848;
    i1848 = 4096 * i1847;
    int i2162;
    i2162 = 32 * i1846;
    int i2656;
    i2656 = 8192 * i1847;
    int i3259;
    i3259 = i161 % 4;
    int i3313;
    i3313 = 8192 * i3259;
    int i3334;
    i3334 = i3333 + i3313;
    int i3527;
    i3527 = 4096 * i3259;
    int i3528;
    i3528 = i3526 + i3527;
    int i3743;
    i3743 = i3311 + i3313;
    int i4054;
    i4054 = i3501 + i3527;
    #pragma unroll
    for(nvfuser_index_t i154 = 0; i154 < 2; ++i154) {
      int i1915;
      i1915 = i414 + (2048 * i154);
      int i1916;
      i1916 = i1915 / 128;
      int i1919;
      i1919 = i1916 % 8;
      int i1923;
      i1923 = i1915 % 128;
      int i1924;
      i1924 = i1923 / 8;
      int i7288;
      i7288 = i5077 + (2048 * (i154 + nvfuser_zero));
      int i7289;
      i7289 = i7288 % 128;
      int i7290;
      i7290 = i7289 / 8;
      int i7300;
      i7300 = i7288 / 128;
      bool b7307;
      b7307 = ((((i644 + (64 * (i7290 / 8))) + (8 * (i7290 % 8))) + (i7289 % 8)) < T1.size[1]) && (((i2162 + (8 * (i7300 / 8))) + (i7300 % 8)) < T0.size[0]);
      if ((!b7307)) {
        arraySet<__half, 8>(&T7[((((i1848 + (128 * ((8 * (i1916 / 8)) + i1919))) + (64 * (i1924 / 8))) + (8 * (i1919 ^ (i1924 % 8)))) + (i1923 % 8))], (__half)0);
      }
    }
    NVFUSER_UPDATE_MAGIC_ZERO
    #pragma unroll
    for(nvfuser_index_t i154 = 0; i154 < 2; ++i154) {
      int i2097;
      i2097 = 2048 * (i154 + nvfuser_zero);
      int i2151;
      i2151 = i414 + i2097;
      int i2152;
      i2152 = i2151 % 128;
      int i2153;
      i2153 = i2152 / 8;
      int i2163;
      i2163 = i2151 / 128;
      int i2404;
      i2404 = i414 + (2048 * i154);
      int i2405;
      i2405 = i2404 / 128;
      int i2408;
      i2408 = i2405 % 8;
      int i2412;
      i2412 = i2404 % 128;
      int i2413;
      i2413 = i2412 / 8;
      int i7917;
      i7917 = i5077 + i2097;
      int i7918;
      i7918 = i7917 % 128;
      int i7919;
      i7919 = i7918 / 8;
      int i7929;
      i7929 = i7917 / 128;
      Ampere::cpAsync(reinterpret_cast<Array<__half,8,8>*>(&T7[((((i1848 + (128 * ((8 * (i2405 / 8)) + i2408))) + (64 * (i2413 / 8))) + (8 * (i2408 ^ (i2413 % 8)))) + (i2412 % 8))]),reinterpret_cast<Array<__half,8,8>*>(&T1[((((i644 + (64 * (i2153 / 8))) + (8 * (i2153 % 8))) + (i2152 % 8)) + (T1.size[1] * ((i2162 + (8 * (i2163 / 8))) + (i2163 % 8))))]),(((((i644 + (64 * (i7919 / 8))) + (8 * (i7919 % 8))) + (i7918 % 8)) < T1.size[1]) && (((i2162 + (8 * (i7929 / 8))) + (i7929 % 8)) < T0.size[0])));
    }
    NVFUSER_UPDATE_MAGIC_ZERO
    #pragma unroll
    for(nvfuser_index_t i143 = 0; i143 < 4; ++i143) {
      int i2657;
      i2657 = i414 + (2048 * i143);
      int i2658;
      i2658 = i2657 / 256;
      int i2661;
      i2661 = i2658 % 8;
      int i2665;
      i2665 = i2657 % 256;
      int i2666;
      i2666 = i2665 / 8;
      int i8410;
      i8410 = i5077 + (2048 * (i143 + nvfuser_zero));
      int i8411;
      i8411 = i8410 % 256;
      int i8412;
      i8412 = i8411 / 8;
      int i8422;
      i8422 = i8410 / 256;
      bool b8429;
      b8429 = ((((i1361 + (64 * (i8412 / 8))) + (8 * (i8412 % 8))) + (i8411 % 8)) < T0.size[1]) && (((i2162 + (8 * (i8422 / 8))) + (i8422 % 8)) < T0.size[0]);
      if ((!b8429)) {
        arraySet<__half, 8>(&T6[((((i2656 + (256 * ((8 * (i2658 / 8)) + i2661))) + (64 * (i2666 / 8))) + (8 * (i2661 ^ (i2666 % 8)))) + (i2665 % 8))], (__half)0);
      }
    }
    NVFUSER_UPDATE_MAGIC_ZERO
    #pragma unroll
    for(nvfuser_index_t i143 = 0; i143 < 4; ++i143) {
      int i2839;
      i2839 = 2048 * (i143 + nvfuser_zero);
      int i2893;
      i2893 = i414 + i2839;
      int i2894;
      i2894 = i2893 % 256;
      int i2895;
      i2895 = i2894 / 8;
      int i2904;
      i2904 = i2893 / 256;
      int i3145;
      i3145 = i414 + (2048 * i143);
      int i3146;
      i3146 = i3145 / 256;
      int i3149;
      i3149 = i3146 % 8;
      int i3153;
      i3153 = i3145 % 256;
      int i3154;
      i3154 = i3153 / 8;
      int i9039;
      i9039 = i5077 + i2839;
      int i9040;
      i9040 = i9039 % 256;
      int i9041;
      i9041 = i9040 / 8;
      int i9051;
      i9051 = i9039 / 256;
      Ampere::cpAsync(reinterpret_cast<Array<__half,8,8>*>(&T6[((((i2656 + (256 * ((8 * (i3146 / 8)) + i3149))) + (64 * (i3154 / 8))) + (8 * (i3149 ^ (i3154 % 8)))) + (i3153 % 8))]),reinterpret_cast<Array<__half,8,8>*>(&T0[((((i1361 + (64 * (i2895 / 8))) + (8 * (i2895 % 8))) + (i2894 % 8)) + (T0.size[1] * ((i2162 + (8 * (i2904 / 8))) + (i2904 % 8))))]),(((((i1361 + (64 * (i9041 / 8))) + (8 * (i9041 % 8))) + (i9040 % 8)) < T0.size[1]) && (((i2162 + (8 * (i9051 / 8))) + (i9051 % 8)) < T0.size[0])));
    }
    NVFUSER_UPDATE_MAGIC_ZERO
    Array<__half, (((ceilDiv(64, 16)) * 8) * 2), 8> T8;
    Array<__half, (((ceilDiv(64, 16)) * 8) * 2), 8> T9;
    #pragma unroll
    for(nvfuser_index_t i145 = 0; i145 < 4; ++i145) {
      int i3320;
      i3320 = i3318 + (16 * i145);
      Turing::ldMatrixT (*reinterpret_cast<Array<__half,8,8>*>(&T8[(8 * i145)]),&T6[((i3334 + (8 * (i3330 ^ (i3320 / 8)))) + (i3320 % 8))]);
    }
    NVFUSER_UPDATE_MAGIC_ZERO
    #pragma unroll
    for(nvfuser_index_t i156 = 0; i156 < 4; ++i156) {
      int i3529;
      i3529 = i3302 + (16 * i156);
      Turing::ldMatrixT (*reinterpret_cast<Array<__half,8,8>*>(&T9[(8 * i156)]),&T7[((i3528 + (8 * (i3523 ^ (i3529 / 8)))) + (i3529 % 8))]);
    }
    NVFUSER_UPDATE_MAGIC_ZERO
    #pragma unroll
    for(nvfuser_index_t i164 = 0; i164 < 1; ++i164) {
      int i3711;
      i3711 = 1 + i164;
      int i3712;
      i3712 = 16 * i3711;
      int i3744;
      i3744 = i3303 + i3712;
      int i3747;
      i3747 = i3744 % 8;
      int i3750;
      i3750 = i3743 + (256 * ((8 * (i3744 / 8)) + i3747));
      int i3789;
      i3789 = 32 * (i3711 % 2);
      int i3855;
      i3855 = 32 * (i164 % 2);
      int i4055;
      i4055 = i3520 + i3712;
      int i4058;
      i4058 = i4055 % 8;
      int i4061;
      i4061 = i4054 + (128 * ((8 * (i4055 / 8)) + i4058));
      #pragma unroll
      for(nvfuser_index_t i145 = 0; i145 < 4; ++i145) {
        int i3751;
        i3751 = i3318 + (16 * i145);
        Turing::ldMatrixT (*reinterpret_cast<Array<__half,8,8>*>(&T8[(i3789 + (8 * i145))]),&T6[((i3750 + (i3751 % 8)) + (8 * (i3747 ^ (i3751 / 8))))]);
      }
      __half T2[((ceilDiv(64, 16)) * 8)];
      #pragma unroll
      for(nvfuser_index_t i147 = 0; i147 < 4; ++i147) {
        int i3808;
        i3808 = 8 * i147;
        #pragma unroll
        for(nvfuser_index_t i149 = 0; i149 < 8; ++i149) {
          T2[(i3808 + i149)] = 0.00000000000000000e+00;
        }
      }
      #pragma unroll
      for(nvfuser_index_t i147 = 0; i147 < 4; ++i147) {
        int i3827;
        i3827 = 8 * i147;
        int i3859;
        i3859 = i3855 + i3827;
        #pragma unroll
        for(nvfuser_index_t i149 = 0; i149 < 8; ++i149) {
          T2[(i3827 + i149)]
             = T8[(i3859 + i149)];
        }
      }
      #pragma unroll
      for(nvfuser_index_t i156 = 0; i156 < 4; ++i156) {
        int i4062;
        i4062 = i3302 + (16 * i156);
        Turing::ldMatrixT (*reinterpret_cast<Array<__half,8,8>*>(&T9[(i3789 + (8 * i156))]),&T7[((i4061 + (i4062 % 8)) + (8 * (i4058 ^ (i4062 / 8))))]);
      }
      __half T3[((ceilDiv(64, 16)) * 8)];
      #pragma unroll
      for(nvfuser_index_t i158 = 0; i158 < 4; ++i158) {
        int i4117;
        i4117 = 8 * i158;
        #pragma unroll
        for(nvfuser_index_t i159 = 0; i159 < 8; ++i159) {
          T3[(i4117 + i159)] = 0.00000000000000000e+00;
        }
      }
      #pragma unroll
      for(nvfuser_index_t i158 = 0; i158 < 4; ++i158) {
        int i4136;
        i4136 = 8 * i158;
        int i4168;
        i4168 = i3855 + i4136;
        #pragma unroll
        for(nvfuser_index_t i159 = 0; i159 < 8; ++i159) {
          T3[(i4136 + i159)]
             = T9[(i4168 + i159)];
        }
      }
      #pragma unroll
      for(nvfuser_index_t i166 = 0; i166 < 4; ++i166) {
        int i4226;
        i4226 = 32 * i166;
        #pragma unroll
        for(nvfuser_index_t i167 = 0; i167 < 4; ++i167) {
          Ampere::M16N16K16TN<16>(
            reinterpret_cast<Array<float,8,8>*>(&T5[(i4226 + (4 * i167))]),
            &(reinterpret_cast<Array<__half,8,8>*>(&T2)[i166]),
            &(reinterpret_cast<Array<__half,8,8>*>(&T3)[i167]));
        }
      }
    }
    NVFUSER_UPDATE_MAGIC_ZERO
    __half T2[((ceilDiv(64, 16)) * 8)];
    #pragma unroll
    for(nvfuser_index_t i147 = 0; i147 < 4; ++i147) {
      int i4246;
      i4246 = 8 * i147;
      #pragma unroll
      for(nvfuser_index_t i149 = 0; i149 < 8; ++i149) {
        T2[(i4246 + i149)] = 0.00000000000000000e+00;
      }
    }
    #pragma unroll
    for(nvfuser_index_t i147 = 0; i147 < 4; ++i147) {
      int i4265;
      i4265 = 8 * i147;
      int i4295;
      i4295 = 32 + i4265;
      #pragma unroll
      for(nvfuser_index_t i149 = 0; i149 < 8; ++i149) {
        T2[(i4265 + i149)]
           = T8[(i4295 + i149)];
      }
    }
    __half T3[((ceilDiv(64, 16)) * 8)];
    #pragma unroll
    for(nvfuser_index_t i158 = 0; i158 < 4; ++i158) {
      int i4314;
      i4314 = 8 * i158;
      #pragma unroll
      for(nvfuser_index_t i159 = 0; i159 < 8; ++i159) {
        T3[(i4314 + i159)] = 0.00000000000000000e+00;
      }
    }
    #pragma unroll
    for(nvfuser_index_t i158 = 0; i158 < 4; ++i158) {
      int i4333;
      i4333 = 8 * i158;
      int i4363;
      i4363 = 32 + i4333;
      #pragma unroll
      for(nvfuser_index_t i159 = 0; i159 < 8; ++i159) {
        T3[(i4333 + i159)]
           = T9[(i4363 + i159)];
      }
    }
    #pragma unroll
    for(nvfuser_index_t i166 = 0; i166 < 4; ++i166) {
      int i4421;
      i4421 = 32 * i166;
      #pragma unroll
      for(nvfuser_index_t i167 = 0; i167 < 4; ++i167) {
        Ampere::M16N16K16TN<16>(
          reinterpret_cast<Array<float,8,8>*>(&T5[(i4421 + (4 * i167))]),
          &(reinterpret_cast<Array<__half,8,8>*>(&T2)[i166]),
          &(reinterpret_cast<Array<__half,8,8>*>(&T3)[i167]));
      }
    }
    NVFUSER_UPDATE_MAGIC_ZERO
    Ampere::cpAsyncPartialBarrier<2>();
    __barrier_sync(0);
    Ampere::cpAsyncCommit();
  }
  #pragma unroll
  for(nvfuser_index_t i172 = 0; i172 < 4; ++i172) {
    int i4475;
    i4475 = 2 * i172;
    int i4609;
    i4609 = i4608 + (16 * i172);
    #pragma unroll
    for(nvfuser_index_t i173 = 0; i173 < 4; ++i173) {
      int i4472;
      i4472 = 4 * i173;
      int i4605;
      i4605 = i4604 + (16 * i173);
      #pragma unroll
      for(nvfuser_index_t i174 = 0; i174 < 2; ++i174) {
        int i4474;
        i4474 = i4472 + (2 * i174);
        int i4606;
        i4606 = i4605 + (8 * i174);
        #pragma unroll
        for(nvfuser_index_t i175 = 0; i175 < 2; ++i175) {
          int i4478;
          i4478 = i4474 + (16 * (i4475 + i175));
          int i4612;
          i4612 = i4606 + (T1.size[1] * (i4609 + (8 * i175)));
          bool b9214;
          b9214 = (i4609 + (8 * (i175 + nvfuser_zero))) < T0.size[1];
          #pragma unroll
          for(nvfuser_index_t i176 = 0; i176 < 2; ++i176) {
            int i4480;
            i4480 = i176 + nvfuser_zero;
            if ((b9214 && ((i4606 + i4480) < T1.size[1]))) {
              T4[(i4612 + i4480)]
                 = T5[(i4478 + i176)];
            }
          }
        }
      }
    }
  }
  NVFUSER_UPDATE_MAGIC_ZERO
}

Geänderter Text

Datei öffnen

__global__ void kernel4(Tensor<__half, 2> T0, Tensor<__half, 2> T1, Tensor<float, 2> T4) {
  alignas(16) extern __shared__ char array[];
  unsigned smem_offset = 0;
  NVFUSER_DEFINE_MAGIC_ZERO
  int i1864;
  i1864 = (((nvfuser_index_t)threadIdx.z) * 4) + (2 * ((nvfuser_index_t)threadIdx.y));
  int i1865;
  i1865 = i1864 + (((nvfuser_index_t)threadIdx.x) / 16);
  int i1889;
  i1889 = i1865 % 8;
  int i1229;
  i1229 = ((nvfuser_index_t)threadIdx.x) % 16;
  int i1866;
  i1866 = i1865 / 8;
  int i1875;
  i1875 = (i1229 / 8) * 64;
  int i1893;
  i1893 = i1229 % 8;
  int i1896;
  i1896 = (((1024 * i1866) + (128 * i1889)) + i1875) + ((i1889 ^ i1893) * 8);
  int i3166;
  i3166 = T1.size[1] * 8;
  int i3192;
  i3192 = i3166 * i1866;
  int i3193;
  i3193 = T1.size[1] * i1889;
  int i3180;
  i3180 = ((nvfuser_index_t)blockIdx.y) * 128;
  int i3197;
  i3197 = i1893 * 8;
  int i3198;
  i3198 = (((i3192 + i3193) + i3180) + i1875) + i3197;
  int i3186;
  i3186 = T1.size[1] * 32;
  int i3189;
  i3189 = T1.size[1] * 16;
  int i6198;
  i6198 = (((nvfuser_index_t)threadIdx.z) * 2) + ((nvfuser_index_t)threadIdx.y);
  int i6199;
  i6199 = i6198 + (((nvfuser_index_t)threadIdx.x) / 32);
  int i6221;
  i6221 = i6199 % 8;
  int i5800;
  i5800 = ((nvfuser_index_t)threadIdx.x) % 32;
  int i6200;
  i6200 = i6199 / 8;
  int i6208;
  i6208 = (i5800 / 8) * 64;
  int i6225;
  i6225 = i5800 % 8;
  int i6228;
  i6228 = (((2048 * i6200) + (256 * i6221)) + i6208) + ((i6221 ^ i6225) * 8);
  int i7323;
  i7323 = T0.size[1] * 8;
  int i7347;
  i7347 = i7323 * i6200;
  int i7348;
  i7348 = T0.size[1] * i6221;
  int i7335;
  i7335 = ((nvfuser_index_t)blockIdx.x) * 256;
  int i7352;
  i7352 = i6225 * 8;
  int i7353;
  i7353 = (((i7347 + i7348) + i7335) + i6208) + i7352;
  int i7341;
  i7341 = T0.size[1] * 32;
  int i11742;
  i11742 = (((((T1.size[1] * 96) + i3192) + i3193) + i3180) + i1875) + i3197;
  int i16008;
  i16008 = (((((T0.size[1] * 96) + i7347) + i7348) + i7335) + i6208) + i7352;
  int i17482;
  i17482 = ((nvfuser_index_t)threadIdx.x) % 8;
  int i17724;
  i17724 = i17482 % 8;
  int i17481;
  i17481 = ((nvfuser_index_t)threadIdx.x) / 8;
  int i17715;
  i17715 = i17482 / 8;
  int i17550;
  i17550 = i17481 / 2;
  int i17829;
  i17829 = 2048 * i17550;
  int i17830;
  i17830 = 2048 * i17715;
  int i17832;
  i17832 = 256 * i17724;
  int i17834;
  i17834 = ((nvfuser_index_t)threadIdx.z) * 64;
  int i17835;
  i17835 = ((i17829 + i17830) + i17832) + i17834;
  int i17570;
  i17570 = i17481 % 2;
  int i18255;
  i18255 = 1024 * i17570;
  int i18256;
  i18256 = 1024 * i17715;
  int i18258;
  i18258 = 128 * i17724;
  int i18246;
  i18246 = ((nvfuser_index_t)threadIdx.y) * 64;
  int i18260;
  i18260 = ((i18255 + i18256) + i18258) + i18246;
  int i18918;
  i18918 = (((4096 + i17829) + i17830) + i17832) + i17834;
  int i19677;
  i19677 = (((2048 + i18255) + i18256) + i18258) + i18246;
  int i20235;
  i20235 = ((nvfuser_index_t)threadIdx.x) / 4;
  int i20286;
  i20286 = (((nvfuser_index_t)threadIdx.x) % 4) * 2;
  int i20301;
  i20301 = ((((((T1.size[1] * ((nvfuser_index_t)blockIdx.x)) * 256) + ((T1.size[1] * ((nvfuser_index_t)threadIdx.z)) * 64)) + (T1.size[1] * i20235)) + i3180) + i18246) + i20286;
  int i22576;
  i22576 = (8 * ((nvfuser_index_t)threadIdx.x)) + 7;
  int i22577;
  i22577 = i22576 % 128;
  int i22578;
  i22578 = i22577 / 8;
  bool b22625;
  b22625 = (((i3180 + ((i22578 / 8) * 64)) + ((i22578 % 8) * 8)) + (i22577 % 8)) < T1.size[1];
  int i22627;
  i22627 = i1864 + (i22576 / 128);
  int i22629;
  i22629 = 8 * (i22627 / 8);
  int i22630;
  i22630 = i22627 % 8;
  int i22631;
  i22631 = i22629 + i22630;
  int i27301;
  i27301 = i22576 % 256;
  int i27302;
  i27302 = i27301 / 8;
  bool b27311;
  b27311 = (((i7335 + ((i27302 / 8) * 64)) + ((i27302 % 8) * 8)) + (i27301 % 8)) < T0.size[1];
  int i27313;
  i27313 = i6198 + (i22576 / 256);
  int i27315;
  i27315 = 8 * (i27313 / 8);
  int i27316;
  i27316 = i27313 % 8;
  int i27317;
  i27317 = i27315 + i27316;
  int i31979;
  i31979 = (96 + i22629) + i22630;
  int i36682;
  i36682 = (96 + i27315) + i27316;
  int i39223;
  i39223 = (i7335 + i17834) + i20235;
  int i39228;
  i39228 = (i3180 + i18246) + i20286;
  smem_offset = alignBufferSize(smem_offset, 16);
  __half* T7 = reinterpret_cast<__half*>(array + smem_offset);
  smem_offset += (((((((ceilDiv(((((ceilDiv(32, 8)) * (ceilDiv(8, 1))) * 1) * (((ceilDiv(128, 64)) * (ceilDiv(64, 8))) * 8)), 2048)) * (ceilDiv((ceilDiv((ceilDiv(2048, 8)), 32)), 2))) * 2) * 32) * 8) * 4) * sizeof(__half));
  smem_offset = alignBufferSize(smem_offset, 16);
  __half* T6 = reinterpret_cast<__half*>(array + smem_offset);
  smem_offset += (((((((ceilDiv(((((ceilDiv(32, 8)) * (ceilDiv(8, 1))) * 1) * (((ceilDiv(256, 64)) * (ceilDiv(64, 8))) * 8)), 2048)) * (ceilDiv((ceilDiv((ceilDiv(2048, 8)), 32)), 2))) * 2) * 32) * 8) * 4) * sizeof(__half));
  float T5[(((((ceilDiv(64, 16)) * (ceilDiv(64, 16))) * (ceilDiv(16, 8))) * (ceilDiv(16, 8))) * 2)];
  #pragma unroll
  for(nvfuser_index_t i166 = 0; i166 < 4; ++i166) {
    int i251;
    i251 = 32 * i166;
    #pragma unroll
    for(nvfuser_index_t i167 = 0; i167 < 4; ++i167) {
      Ampere::initM16N16K16TN<16>(reinterpret_cast<Array<float,8,8>*>(&T5[(i251 + (4 * i167))]));
    }
  }
  NVFUSER_UPDATE_MAGIC_ZERO
  #pragma unroll
  for(nvfuser_index_t i160 = 0; i160 < 3; ++i160) {
    int i1897;
    i1897 = i1896 + (4096 * i160);
    int i3199;
    i3199 = i3198 + (i3186 * i160);
    int i6229;
    i6229 = i6228 + (8192 * i160);
    int i7354;
    i7354 = i7353 + (i7341 * i160);
    int i22613;
    i22613 = 32 * i160;
    int i22632;
    i22632 = i22631 + i22613;
    int i27318;
    i27318 = i27317 + i22613;
    #pragma unroll
    for(nvfuser_index_t i154 = 0; i154 < 2; ++i154) {
      bool b22635;
      b22635 = b22625 && ((i22632 + (16 * (i154 + nvfuser_zero))) < T0.size[0]);
      if ((!b22635)) {
        arraySet<__half, 8>(&T7[(i1897 + (2048 * i154))], (__half)0);
      }
    }
    #pragma unroll
    for(nvfuser_index_t i154 = 0; i154 < 2; ++i154) {
      int i1912;
      i1912 = i154 + nvfuser_zero;
      Ampere::cpAsync(reinterpret_cast<Array<__half,8,8>*>(&T7[(i1897 + (2048 * i154))]),reinterpret_cast<Array<__half,8,8>*>(&T1[(i3199 + (i3189 * i1912))]),(b22625 && ((i22632 + (16 * i1912)) < T0.size[0])));
    }
    #pragma unroll
    for(nvfuser_index_t i143 = 0; i143 < 4; ++i143) {
      bool b27321;
      b27321 = b27311 && ((i27318 + (8 * (i143 + nvfuser_zero))) < T0.size[0]);
      if ((!b27321)) {
        arraySet<__half, 8>(&T6[(i6229 + (2048 * i143))], (__half)0);
      }
    }
    #pragma unroll
    for(nvfuser_index_t i143 = 0; i143 < 4; ++i143) {
      int i6244;
      i6244 = i143 + nvfuser_zero;
      Ampere::cpAsync(reinterpret_cast<Array<__half,8,8>*>(&T6[(i6229 + (2048 * i143))]),reinterpret_cast<Array<__half,8,8>*>(&T0[(i7354 + (i7323 * i6244))]),(b27311 && ((i27318 + (8 * i6244)) < T0.size[0])));
    }
    Ampere::cpAsyncCommit();
  }
  NVFUSER_UPDATE_MAGIC_ZERO
  Ampere::cpAsyncPartialBarrier<2>();
  __barrier_sync(0);
  #pragma unroll 1
  for(nvfuser_index_t i161 = 0; i161 < (ceilDiv(T0.size[0], 32)); ++i161) {
    int i10426;
    i10426 = (3 + i161) % 4;
    int i10431;
    i10431 = i1896 + (4096 * i10426);
    int i11744;
    i11744 = i11742 + (i3186 * i161);
    int i14873;
    i14873 = i6228 + (8192 * i10426);
    int i16010;
    i16010 = i16008 + (i7341 * i161);
    int i17544;
    i17544 = i161 % 4;
    int i17836;
    i17836 = 8192 * i17544;
    int i17837;
    i17837 = i17835 + i17836;
    int i18261;
    i18261 = 4096 * i17544;
    int i18262;
    i18262 = i18260 + i18261;
    int i18919;
    i18919 = i18918 + i17836;
    int i19678;
    i19678 = i19677 + i18261;
    int i31972;
    i31972 = 32 * i161;
    int i31980;
    i31980 = i31979 + i31972;
    int i36683;
    i36683 = i36682 + i31972;
    #pragma unroll
    for(nvfuser_index_t i154 = 0; i154 < 2; ++i154) {
      bool b31983;
      b31983 = b22625 && ((i31980 + (16 * (i154 + nvfuser_zero))) < T0.size[0]);
      if ((!b31983)) {
        arraySet<__half, 8>(&T7[(i10431 + (2048 * i154))], (__half)0);
      }
    }
    NVFUSER_UPDATE_MAGIC_ZERO
    #pragma unroll
    for(nvfuser_index_t i154 = 0; i154 < 2; ++i154) {
      int i10446;
      i10446 = i154 + nvfuser_zero;
      Ampere::cpAsync(reinterpret_cast<Array<__half,8,8>*>(&T7[(i10431 + (2048 * i154))]),reinterpret_cast<Array<__half,8,8>*>(&T1[(i11744 + (i3189 * i10446))]),(b22625 && ((i31980 + (16 * i10446)) < T0.size[0])));
    }
    NVFUSER_UPDATE_MAGIC_ZERO
    #pragma unroll
    for(nvfuser_index_t i143 = 0; i143 < 4; ++i143) {
      bool b36686;
      b36686 = b27311 && ((i36683 + (8 * (i143 + nvfuser_zero))) < T0.size[0]);
      if ((!b36686)) {
        arraySet<__half, 8>(&T6[(i14873 + (2048 * i143))], (__half)0);
      }
    }
    NVFUSER_UPDATE_MAGIC_ZERO
    #pragma unroll
    for(nvfuser_index_t i143 = 0; i143 < 4; ++i143) {
      int i14888;
      i14888 = i143 + nvfuser_zero;
      Ampere::cpAsync(reinterpret_cast<Array<__half,8,8>*>(&T6[(i14873 + (2048 * i143))]),reinterpret_cast<Array<__half,8,8>*>(&T0[(i16010 + (i7323 * i14888))]),(b27311 && ((i36683 + (8 * i14888)) < T0.size[0])));
    }
    NVFUSER_UPDATE_MAGIC_ZERO
    Array<__half, (((ceilDiv(64, 16)) * 8) * 2), 8> T8;
    Array<__half, (((ceilDiv(64, 16)) * 8) * 2), 8> T9;
    #pragma unroll
    for(nvfuser_index_t i145 = 0; i145 < 4; ++i145) {
      Turing::ldMatrixT (*reinterpret_cast<Array<__half,8,8>*>(&T8[(8 * i145)]),&T6[(i17837 + (8 * (i17724 ^ (i17570 + (2 * i145)))))]);
    }
    NVFUSER_UPDATE_MAGIC_ZERO
    #pragma unroll
    for(nvfuser_index_t i156 = 0; i156 < 4; ++i156) {
      Turing::ldMatrixT (*reinterpret_cast<Array<__half,8,8>*>(&T9[(8 * i156)]),&T7[(i18262 + (8 * (i17724 ^ (i17550 + (2 * i156)))))]);
    }
    NVFUSER_UPDATE_MAGIC_ZERO
    #pragma unroll
    for(nvfuser_index_t i164 = 0; i164 < 1; ++i164) {
      int i18920;
      i18920 = i18919 + (4096 * i164);
      int i18960;
      i18960 = 32 * ((1 + i164) % 2);
      int i19035;
      i19035 = 32 * (i164 % 2);
      int i19679;
      i19679 = i19678 + (2048 * i164);
      #pragma unroll
      for(nvfuser_index_t i145 = 0; i145 < 4; ++i145) {
        Turing::ldMatrixT (*reinterpret_cast<Array<__half,8,8>*>(&T8[(i18960 + (8 * i145))]),&T6[(i18920 + (8 * (i17724 ^ (i17570 + (2 * i145)))))]);
      }
      __half T2[((ceilDiv(64, 16)) * 8)];
      #pragma unroll
      for(nvfuser_index_t i147 = 0; i147 < 4; ++i147) {
        int i18982;
        i18982 = 8 * i147;
        #pragma unroll
        for(nvfuser_index_t i149 = 0; i149 < 8; ++i149) {
          T2[(i18982 + i149)] = 0.00000000000000000e+00;
        }
      }
      #pragma unroll
      for(nvfuser_index_t i147 = 0; i147 < 4; ++i147) {
        int i19003;
        i19003 = 8 * i147;
        int i19039;
        i19039 = i19035 + i19003;
        #pragma unroll
        for(nvfuser_index_t i149 = 0; i149 < 8; ++i149) {
          T2[(i19003 + i149)]
             = T8[(i19039 + i149)];
        }
      }
      #pragma unroll
      for(nvfuser_index_t i156 = 0; i156 < 4; ++i156) {
        Turing::ldMatrixT (*reinterpret_cast<Array<__half,8,8>*>(&T9[(i18960 + (8 * i156))]),&T7[(i19679 + (8 * (i17724 ^ (i17550 + (2 * i156)))))]);
      }
      __half T3[((ceilDiv(64, 16)) * 8)];
      #pragma unroll
      for(nvfuser_index_t i158 = 0; i158 < 4; ++i158) {
        int i19742;
        i19742 = 8 * i158;
        #pragma unroll
        for(nvfuser_index_t i159 = 0; i159 < 8; ++i159) {
          T3[(i19742 + i159)] = 0.00000000000000000e+00;
        }
      }
      #pragma unroll
      for(nvfuser_index_t i158 = 0; i158 < 4; ++i158) {
        int i19763;
        i19763 = 8 * i158;
        int i19799;
        i19799 = i19035 + i19763;
        #pragma unroll
        for(nvfuser_index_t i159 = 0; i159 < 8; ++i159) {
          T3[(i19763 + i159)]
             = T9[(i19799 + i159)];
        }
      }
      #pragma unroll
      for(nvfuser_index_t i166 = 0; i166 < 4; ++i166) {
        int i19866;
        i19866 = 32 * i166;
        #pragma unroll
        for(nvfuser_index_t i167 = 0; i167 < 4; ++i167) {
          Ampere::M16N16K16TN<16>(
            reinterpret_cast<Array<float,8,8>*>(&T5[(i19866 + (4 * i167))]),
            &(reinterpret_cast<Array<__half,8,8>*>(&T2)[i166]),
            &(reinterpret_cast<Array<__half,8,8>*>(&T3)[i167]));
        }
      }
    }
    NVFUSER_UPDATE_MAGIC_ZERO
    __half T2[((ceilDiv(64, 16)) * 8)];
    #pragma unroll
    for(nvfuser_index_t i147 = 0; i147 < 4; ++i147) {
      int i19887;
      i19887 = 8 * i147;
      #pragma unroll
      for(nvfuser_index_t i149 = 0; i149 < 8; ++i149) {
        T2[(i19887 + i149)] = 0.00000000000000000e+00;
      }
    }
    #pragma unroll
    for(nvfuser_index_t i147 = 0; i147 < 4; ++i147) {
      int i19907;
      i19907 = 8 * i147;
      int i19938;
      i19938 = 32 + i19907;
      #pragma unroll
      for(nvfuser_index_t i149 = 0; i149 < 8; ++i149) {
        T2[(i19907 + i149)]
           = T8[(i19938 + i149)];
      }
    }
    __half T3[((ceilDiv(64, 16)) * 8)];
    #pragma unroll
    for(nvfuser_index_t i158 = 0; i158 < 4; ++i158) {
      int i19958;
      i19958 = 8 * i158;
      #pragma unroll
      for(nvfuser_index_t i159 = 0; i159 < 8; ++i159) {
        T3[(i19958 + i159)] = 0.00000000000000000e+00;
      }
    }
    #pragma unroll
    for(nvfuser_index_t i158 = 0; i158 < 4; ++i158) {
      int i19978;
      i19978 = 8 * i158;
      int i20009;
      i20009 = 32 + i19978;
      #pragma unroll
      for(nvfuser_index_t i159 = 0; i159 < 8; ++i159) {
        T3[(i19978 + i159)]
           = T9[(i20009 + i159)];
      }
    }
    #pragma unroll
    for(nvfuser_index_t i166 = 0; i166 < 4; ++i166) {
      int i20073;
      i20073 = 32 * i166;
      #pragma unroll
      for(nvfuser_index_t i167 = 0; i167 < 4; ++i167) {
        Ampere::M16N16K16TN<16>(
          reinterpret_cast<Array<float,8,8>*>(&T5[(i20073 + (4 * i167))]),
          &(reinterpret_cast<Array<__half,8,8>*>(&T2)[i166]),
          &(reinterpret_cast<Array<__half,8,8>*>(&T3)[i167]));
      }
    }
    NVFUSER_UPDATE_MAGIC_ZERO
    Ampere::cpAsyncPartialBarrier<2>();
    __barrier_sync(0);
    Ampere::cpAsyncCommit();
  }
  #pragma unroll
  for(nvfuser_index_t i172 = 0; i172 < 4; ++i172) {
    int i20139;
    i20139 = 32 * i172;
    int i20303;
    i20303 = i20301 + (i3189 * i172);
    int i39224;
    i39224 = i39223 + (16 * i172);
    #pragma unroll
    for(nvfuser_index_t i173 = 0; i173 < 4; ++i173) {
      int i20141;
      i20141 = i20139 + (4 * i173);
      int i20291;
      i20291 = 16 * i173;
      int i20304;
      i20304 = i20303 + i20291;
      int i39229;
      i39229 = i39228 + i20291;
      #pragma unroll
      for(nvfuser_index_t i174 = 0; i174 < 2; ++i174) {
        int i20143;
        i20143 = i20141 + (2 * i174);
        int i20293;
        i20293 = 8 * i174;
        int i20305;
        i20305 = i20304 + i20293;
        int i39230;
        i39230 = i39229 + i20293;
        #pragma unroll
        for(nvfuser_index_t i175 = 0; i175 < 2; ++i175) {
          int i20145;
          i20145 = i20143 + (16 * i175);
          int i20307;
          i20307 = i20305 + (i3166 * i175);
          bool b39226;
          b39226 = (i39224 + (8 * (i175 + nvfuser_zero))) < T0.size[1];
          #pragma unroll
          for(nvfuser_index_t i176 = 0; i176 < 2; ++i176) {
            int i20147;
            i20147 = i176 + nvfuser_zero;
            if ((b39226 && ((i39230 + i20147) < T1.size[1]))) {
              T4[(i20307 + i20147)]
                 = T5[(i20145 + i176)];
            }
          }
        }
      }
    }
  }
  NVFUSER_UPDATE_MAGIC_ZERO
}