biondizzle/nvfp4-megamoe-kernel
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

P4: use proper CUDA enum names

Browse Source
This commit is contained in:
biondizzle
2026-05-30 08:34:19 +00:00
parent 6d624a1b14
commit 277689f8b8
1 changed files with 36 additions and 81 deletions
Download Patch File Download Diff File Expand all files Collapse all files

117
tests/unit/test_p4_tma_descriptor_dump.cu
View File

@@ -1,6 +1,5 @@
/**
* P4: Dump TMA descriptor bytes for comparison.
* Uses CUDA Driver API directly.
*/
#include <cuda.h>
#include <cuda_runtime.h>
@@ -8,14 +7,6 @@
#include <cstdint>
#include <cstring>
/* CUDA 13.2 enum values for cuTensorMapEncodeTiled:
* CUtensorMapDataType: 0=F16, 6=BF16
* CUtensorMapInterleave: 0=NONE, 1=16B, 2=32B
* CUtensorMapSwizzle: 0=NONE, 1=4B, 2=32B, 3=64B, 4=128B
* CUtensorMapL2Promotion: 0=NONE, 1=64B, 2=128B, 3=256B
* CUtensorMapOOBFill: 0=NONE, 1=ZERO
*/
int main() {
const int ROWS = 128;
const int COLS = 16;
@@ -33,110 +24,74 @@ int main() {
CUtensorMap tma_desc;
CUresult res;
// Descriptor 1: NO swizzle
res = cuTensorMapEncodeTiled(
&tma_desc,
(CUtensorMapDataType)6, // BF16
2, // rank
d_ptr,
tensorDims,
globalStrides,
boxDims,
elementStrides,
(CUtensorMapInterleave)0, // NONE
(CUtensorMapSwizzle)0, // NONE
0, // NONE
(CUtensorMapOOBFill)0 // NONE
);
// Config 1: NO swizzle
printf("=== Descriptor 1: NO swizzle ===\n");
res = cuTensorMapEncodeTiled(
&tma_desc, 2, CU_TENSOR_MAP_DATA_TYPE_BFLOAT16,
d_ptr, tensorDims, globalStrides, boxDims, elementStrides,
CU_TENSOR_MAP_INTERLEAVE_NONE, CU_TENSOR_MAP_SWIZZLE_NONE,
CU_TENSOR_MAP_L2_PROMOTION_NONE, CU_TENSOR_MAP_OOB_FILL_NONE
);
if (res != CUDA_SUCCESS) { printf("FAILED: %d\n", res); }
else {
auto* bytes = reinterpret_cast<const uint8_t*>(&tma_desc);
auto* b = reinterpret_cast<const uint8_t*>(&tma_desc);
for (int i = 0; i < 128; i += 16) {
printf("[%3d-%3d]: ", i, i+15);
for (int j = 0; j < 16; j++) printf("%02x ", bytes[i+j]);
for (int j = 0; j < 16; j++) printf("%02x ", b[i+j]);
printf("\n");
}
}
// Descriptor 2: SWIZZLE_128B
res = cuTensorMapEncodeTiled(
&tma_desc,
(CUtensorMapDataType)6,
2,
d_ptr,
tensorDims,
globalStrides,
boxDims,
elementStrides,
(CUtensorMapInterleave)0,
(CUtensorMapSwizzle)4, // 128B
0,
(CUtensorMapOOBFill)0
);
// Config 2: SWIZZLE_128B
printf("\n=== Descriptor 2: SWIZZLE_128B ===\n");
res = cuTensorMapEncodeTiled(
&tma_desc, 2, CU_TENSOR_MAP_DATA_TYPE_BFLOAT16,
d_ptr, tensorDims, globalStrides, boxDims, elementStrides,
CU_TENSOR_MAP_INTERLEAVE_NONE, CU_TENSOR_MAP_SWIZZLE_128B,
CU_TENSOR_MAP_L2_PROMOTION_NONE, CU_TENSOR_MAP_OOB_FILL_NONE
);
if (res != CUDA_SUCCESS) { printf("FAILED: %d\n", res); }
else {
auto* bytes = reinterpret_cast<const uint8_t*>(&tma_desc);
auto* b = reinterpret_cast<const uint8_t*>(&tma_desc);
for (int i = 0; i < 128; i += 16) {
printf("[%3d-%3d]: ", i, i+15);
for (int j = 0; j < 16; j++) printf("%02x ", bytes[i+j]);
for (int j = 0; j < 16; j++) printf("%02x ", b[i+j]);
printf("\n");
}
}
// Descriptor 3: NO swizzle, OOB_FILL_ZERO
res = cuTensorMapEncodeTiled(
&tma_desc,
(CUtensorMapDataType)6,
2,
d_ptr,
tensorDims,
globalStrides,
boxDims,
elementStrides,
(CUtensorMapInterleave)0,
(CUtensorMapSwizzle)0,
0,
(CUtensorMapOOBFill)1 // ZERO
);
// Config 3: NO swizzle, OOB_FILL_ZERO
printf("\n=== Descriptor 3: NO swizzle, OOB_FILL_ZERO ===\n");
res = cuTensorMapEncodeTiled(
&tma_desc, 2, CU_TENSOR_MAP_DATA_TYPE_BFLOAT16,
d_ptr, tensorDims, globalStrides, boxDims, elementStrides,
CU_TENSOR_MAP_INTERLEAVE_NONE, CU_TENSOR_MAP_SWIZZLE_NONE,
CU_TENSOR_MAP_L2_PROMOTION_NONE, CU_TENSOR_MAP_OOB_FILL_ZERO
);
if (res != CUDA_SUCCESS) { printf("FAILED: %d\n", res); }
else {
auto* bytes = reinterpret_cast<const uint8_t*>(&tma_desc);
auto* b = reinterpret_cast<const uint8_t*>(&tma_desc);
for (int i = 0; i < 128; i += 16) {
printf("[%3d-%3d]: ", i, i+15);
for (int j = 0; j < 16; j++) printf("%02x ", bytes[i+j]);
for (int j = 0; j < 16; j++) printf("%02x ", b[i+j]);
printf("\n");
}
}
// Descriptor 4: SWIZZLE_128B, OOB_FILL_ZERO
res = cuTensorMapEncodeTiled(
&tma_desc,
(CUtensorMapDataType)6,
2,
d_ptr,
tensorDims,
globalStrides,
boxDims,
elementStrides,
(CUtensorMapInterleave)0,
(CUtensorMapSwizzle)4,
0,
(CUtensorMapOOBFill)1
);
// Config 4: SWIZZLE_128B, OOB_FILL_ZERO
printf("\n=== Descriptor 4: SWIZZLE_128B, OOB_FILL_ZERO ===\n");
res = cuTensorMapEncodeTiled(
&tma_desc, 2, CU_TENSOR_MAP_DATA_TYPE_BFLOAT16,
d_ptr, tensorDims, globalStrides, boxDims, elementStrides,
CU_TENSOR_MAP_INTERLEAVE_NONE, CU_TENSOR_MAP_SWIZZLE_128B,
CU_TENSOR_MAP_L2_PROMOTION_NONE, CU_TENSOR_MAP_OOB_FILL_ZERO
);
if (res != CUDA_SUCCESS) { printf("FAILED: %d\n", res); }
else {
auto* bytes = reinterpret_cast<const uint8_t*>(&tma_desc);
auto* b = reinterpret_cast<const uint8_t*>(&tma_desc);
for (int i = 0; i < 128; i += 16) {
printf("[%3d-%3d]: ", i, i+15);
for (int j = 0; j < 16; j++) printf("%02x ", bytes[i+j]);
for (int j = 0; j < 16; j++) printf("%02x ", b[i+j]);
printf("\n");
}
}