[TL] Add example usage/test case for Dynamic Symbolic #191
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
…ability and maintainability
|
Tail split is required when we apply dynamic symbolic into the last dynamic, otherwise the vector load will be blocked #include <tl_templates/gemm.h>
#include <tl_templates/copy.h>
#include <tl_templates/reduce.h>
#include <tl_templates/ldsm.h>
#include <tl_templates/threadblock_swizzle.h>
extern "C" __global__ void __launch_bounds__(128) main_kernel(half_t* __restrict__ A, half_t* __restrict__ B, half_t* __restrict__ C, int k, int m, int n) {
extern __shared__ __align__(1024) uchar buf_dyn_shmem[];
half_t C_local[32];
#pragma unroll
for (int i = 0; i < 16; ++i) {
*(uint1*)(C_local + (i * 2)) = make_uint1(__pack_half2(half_t(0.000000e+00f), half_t(0.000000e+00f)));
}
if (0 < k) {
#pragma unroll
for (int i_1 = 0; i_1 < 16; ++i_1) {
half_t condval;
if ((((((((int)blockIdx.y) * 64) + (i_1 * 4)) + (((int)threadIdx.x) >> 5)) < m) && ((((int)threadIdx.x) & 31) < k))) {
condval = A[(((((((int)blockIdx.y) * 64) + (i_1 * 4)) + (((int)threadIdx.x) >> 5)) * k) + (((int)threadIdx.x) & 31))];
} else {
condval = half_t(0.000000e+00f);
}
((half_t*)buf_dyn_shmem)[(((((i_1 * 128) + ((((int)threadIdx.x) >> 5) * 32)) + (((((((int)threadIdx.x) & 31) >> 4) + (i_1 & 1)) & 1) * 16)) + ((((((int)threadIdx.x) >> 6) + ((((int)threadIdx.x) & 15) >> 3)) & 1) * 8)) + (((int)threadIdx.x) & 7))] = condval;
}
}
if (0 < k) {
#pragma unroll
for (int i_2 = 0; i_2 < 16; ++i_2) {
half_t condval_1;
if (((((i_2 * 2) + (((int)threadIdx.x) >> 6)) < k) && (((((int)blockIdx.x) * 64) + (((int)threadIdx.x) & 63)) < n))) {
condval_1 = B[(((((int)blockIdx.x) * 64) + (((i_2 * 2) + (((int)threadIdx.x) >> 6)) * n)) + (((int)threadIdx.x) & 63))];
} else {
condval_1 = half_t(0.000000e+00f);
}
((half_t*)buf_dyn_shmem)[(((((((i_2 * 128) + ((((int)threadIdx.x) >> 6) * 64)) + (((((((int)threadIdx.x) & 63) >> 5) + ((i_2 & 3) >> 1)) & 1) * 32)) + (((((((int)threadIdx.x) & 31) >> 4) + (i_2 & 1)) & 1) * 16)) + ((((((int)threadIdx.x) >> 6) + ((((int)threadIdx.x) & 15) >> 3)) & 1) * 8)) + (((int)threadIdx.x) & 7)) + 6144)] = condval_1;
}
tl::cp_async_commit();
}
if (32 < k) {
#pragma unroll
for (int i_3 = 0; i_3 < 16; ++i_3) {
half_t condval_2;
if ((((((((int)blockIdx.y) * 64) + (i_3 * 4)) + (((int)threadIdx.x) >> 5)) < m) && (((((int)threadIdx.x) & 31) + 32) < k))) {
condval_2 = A[((((((((int)blockIdx.y) * 64) + (i_3 * 4)) + (((int)threadIdx.x) >> 5)) * k) + (((int)threadIdx.x) & 31)) + 32)];
} else {
condval_2 = half_t(0.000000e+00f);
}
((half_t*)buf_dyn_shmem)[((((((i_3 * 128) + ((((int)threadIdx.x) >> 5) * 32)) + (((((((int)threadIdx.x) & 31) >> 4) + (i_3 & 1)) & 1) * 16)) + ((((((int)threadIdx.x) >> 6) + ((((int)threadIdx.x) & 15) >> 3)) & 1) * 8)) + (((int)threadIdx.x) & 7)) + 2048)] = condval_2;
}
}
if (32 < k) {
#pragma unroll
for (int i_4 = 0; i_4 < 16; ++i_4) {
half_t condval_3;
if ((((((i_4 * 2) + (((int)threadIdx.x) >> 6)) + 32) < k) && (((((int)blockIdx.x) * 64) + (((int)threadIdx.x) & 63)) < n))) {
condval_3 = B[(((((int)blockIdx.x) * 64) + ((((i_4 * 2) + (((int)threadIdx.x) >> 6)) + 32) * n)) + (((int)threadIdx.x) & 63))];
} else {
condval_3 = half_t(0.000000e+00f);
}
((half_t*)buf_dyn_shmem)[(((((((i_4 * 128) + ((((int)threadIdx.x) >> 6) * 64)) + (((((((int)threadIdx.x) & 63) >> 5) + ((i_4 & 3) >> 1)) & 1) * 32)) + (((((((int)threadIdx.x) & 31) >> 4) + (i_4 & 1)) & 1) * 16)) + ((((((int)threadIdx.x) >> 6) + ((((int)threadIdx.x) & 15) >> 3)) & 1) * 8)) + (((int)threadIdx.x) & 7)) + 8192)] = condval_3;
}
tl::cp_async_commit();
}
for (int k_1 = 0; k_1 < (((k + 31) >> 5) - 2); ++k_1) {
__syncthreads();
#pragma unroll
for (int i_5 = 0; i_5 < 16; ++i_5) {
half_t condval_4;
if ((((((((int)blockIdx.y) * 64) + (i_5 * 4)) + (((int)threadIdx.x) >> 5)) < m) && ((((k_1 * 32) + (((int)threadIdx.x) & 31)) + 64) < k))) {
condval_4 = A[((((k_1 * 32) + ((((((int)blockIdx.y) * 64) + (i_5 * 4)) + (((int)threadIdx.x) >> 5)) * k)) + (((int)threadIdx.x) & 31)) + 64)];
} else {
condval_4 = half_t(0.000000e+00f);
}
((half_t*)buf_dyn_shmem)[((((((((k_1 + 2) % 3) * 2048) + (i_5 * 128)) + ((((int)threadIdx.x) >> 5) * 32)) + (((((((int)threadIdx.x) & 31) >> 4) + (i_5 & 1)) & 1) * 16)) + ((((((int)threadIdx.x) >> 6) + ((((int)threadIdx.x) & 15) >> 3)) & 1) * 8)) + (((int)threadIdx.x) & 7))] = condval_4;
}
#pragma unroll
for (int i_6 = 0; i_6 < 16; ++i_6) {
half_t condval_5;
if (((((((k_1 * 32) + (i_6 * 2)) + (((int)threadIdx.x) >> 6)) + 64) < k) && (((((int)blockIdx.x) * 64) + (((int)threadIdx.x) & 63)) < n))) {
condval_5 = B[((((((int)blockIdx.x) * 64) + (n * 64)) + ((((k_1 * 32) + (i_6 * 2)) + (((int)threadIdx.x) >> 6)) * n)) + (((int)threadIdx.x) & 63))];
} else {
condval_5 = half_t(0.000000e+00f);
}
((half_t*)buf_dyn_shmem)[((((((((((k_1 + 2) % 3) * 2048) + (i_6 * 128)) + ((((int)threadIdx.x) >> 6) * 64)) + (((((((int)threadIdx.x) & 63) >> 5) + ((i_6 & 3) >> 1)) & 1) * 32)) + (((((((int)threadIdx.x) & 31) >> 4) + (i_6 & 1)) & 1) * 16)) + ((((((int)threadIdx.x) >> 6) + ((((int)threadIdx.x) & 15) >> 3)) & 1) * 8)) + (((int)threadIdx.x) & 7)) + 6144)] = condval_5;
}
tl::cp_async_commit();
tl::cp_async_wait<2>();
__syncthreads();
tl::gemm_ss<64, 64, 32, 2, 2, 0, 0>((&(((half_t*)buf_dyn_shmem)[((k_1 % 3) * 2048)])), (&(((half_t*)buf_dyn_shmem)[(((k_1 % 3) * 2048) + 6144)])), (&(C_local[0])));
}
if (33 <= k) {
tl::cp_async_wait<1>();
__syncthreads();
tl::gemm_ss<64, 64, 32, 2, 2, 0, 0>((&(((half_t*)buf_dyn_shmem)[(((((k + 31) >> 5) + 1) % 3) * 2048)])), (&(((half_t*)buf_dyn_shmem)[((((((k + 31) >> 5) + 1) % 3) * 2048) + 6144)])), (&(C_local[0])));
}
if (1 <= k) {
tl::cp_async_wait<0>();
__syncthreads();
tl::gemm_ss<64, 64, 32, 2, 2, 0, 0>((&(((half_t*)buf_dyn_shmem)[(((((k + 31) >> 5) + 2) % 3) * 2048)])), (&(((half_t*)buf_dyn_shmem)[((((((k + 31) >> 5) + 2) % 3) * 2048) + 6144)])), (&(C_local[0])));
}
#pragma unroll
for (int i_7 = 0; i_7 < 32; ++i_7) {
if (((((((((int)blockIdx.y) * 64) + (((i_7 & 7) >> 2) * 32)) + (((((int)threadIdx.x) & 63) >> 5) * 16)) + (((i_7 & 3) >> 1) * 8)) + ((((int)threadIdx.x) & 31) >> 2)) < m) && ((((((((int)blockIdx.x) * 64) + ((i_7 >> 3) * 16)) + ((((int)threadIdx.x) >> 6) * 8)) + ((((int)threadIdx.x) & 3) * 2)) + (i_7 & 1)) < n)) {
C[((((((((int)blockIdx.x) * 64) + ((i_7 >> 3) * 16)) + ((((int)threadIdx.x) >> 6) * 8)) + ((((int)threadIdx.x) & 3) * 2)) + ((((((((int)blockIdx.y) * 64) + (((i_7 & 7) >> 2) * 32)) + (((((int)threadIdx.x) & 63) >> 5) * 16)) + (((i_7 & 3) >> 1) * 8)) + ((((int)threadIdx.x) & 31) >> 2)) * n)) + (i_7 & 1))] = C_local[i_7];
}
}
}
lead to significant drop in performance. |
|
I'm wondering about the meaning of tail split. Is this related to TensorCoreEmitter?
|
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
New Test File:
testing/python/tilelang/test_tilelang_dyanmic_symbolic.pyto test dynamic symbolic operations using theTensorCoreIntrinEmitterclass.Example Usage
Describe Computation with Dynamic Symbolic.
Below is the generated kernel :