DeepSeek/DeepGEMM
1
0
Fork 0
mirror of https://github.com/deepseek-ai/DeepGEMM synced 2025-06-26 23:15:49 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

Grouped GEMM skip useless computation for unaligned Ms

Browse Source
This commit is contained in:
yukuai 2025-05-20 15:31:30 +08:00
parent 391755ada0
commit ccca476ac4
3 changed files with 115 additions and 87 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

154
deep_gemm/include/deep_gemm/fp8_gemm.cuh
View File

@ -271,81 +271,91 @@ fp8_gemm_kernel(float* scales_b, int* grouped_layout,
}
};
// Launch MMAs
launch_k_iterations([&](int k_iter, auto type, auto num_former_iters_type) {
constexpr bool kHasDivisibleStages = std::is_same_v<decltype(type), DivisibleK>;
constexpr int kNumInnerStages = kHasDivisibleStages ? kNumStages : (SHAPE_K % kFullKOfAllStages) / BLOCK_K;
DG_STATIC_ASSERT(kNumInnerStages != 0, "Invalid number of inner stages");
#pragma unroll
for (int s = 0; s < kNumInnerStages; ++ s) {
// Read B scales
float scale_b_0 = ld_shared(smem_scales_b + k_iter * kNumStages + s), scale_b_1;
// NOTES: even some blocks do not need to read the second row, but we still load one to align with other blocks
if constexpr (not kMustUseUniformedScaleB)
scale_b_1 = ld_shared(smem_scales_b + k_iter * kNumStages + s + SHAPE_K_SCALES);
// Wait TMA arrivals
full_barriers[s]->wait((scheduler.current_iter * kNumIterations + k_iter) & 1);
// TODO: remove some useless computation for unaligned Ms
if (!scheduler.is_valid_m(math_wg_idx * WGMMA::M, m_block_idx)) {
// Skip useless computation for unaligned Ms
launch_k_iterations([&](int k_iter, auto type, auto _) {
#pragma unroll
for (uint32_t local_idx = 0; local_idx < BLOCK_M / WAVE_BLOCK_M; ++ local_idx) {
auto m_offset = local_idx * WAVE_BLOCK_M;
// Read A scales
// NOTES: all shared memory read must be prior to `warpgroup_arrive` to avoid next scheduled block polluting the results
auto scale_a_0 = ld_shared(smem_scales_a[s] + r_0 + m_offset);
auto scale_a_1 = ld_shared(smem_scales_a[s] + r_1 + m_offset);
// Commit WGMMA instructions
#pragma unroll
for (int i = 0; i < WGMMA::kNumAccum; ++ i)
warpgroup_fence_operand(accum[i]);
warpgroup_arrive();
#pragma unroll
for (int k = 0; k < BLOCK_K / WGMMA::K; ++ k) {
auto desc_a = make_smem_desc(smem_a[s] + (math_wg_idx * WGMMA::M + m_offset) * BLOCK_K + k * WGMMA::K, 1);
auto desc_b = make_smem_desc(smem_b[s] + k * WGMMA::K, 1);
WGMMA::wgmma(desc_a, desc_b, accum, k);
}
warpgroup_commit_batch();
#pragma unroll
for (int i = 0; i < WGMMA::kNumAccum; ++ i)
warpgroup_fence_operand(accum[i]);
warpgroup_wait<0>();
// Notify barrier arrival at the last warpgroup wave
if (local_idx == BLOCK_M / WAVE_BLOCK_M - 1)
empty_barrier_arrive(s);
// Promote with scales
// NOTES: making it as predicates is very important for performance, comparing to two loops
float scale_0_0 = scale_a_0 * scale_b_0, scale_1_0 = scale_a_1 * scale_b_0;
float scale_0_1, scale_1_1;
if constexpr (not kMustUseUniformedScaleB)
scale_0_1 = scale_a_0 * scale_b_1, scale_1_1 = scale_a_1 * scale_b_1;
auto shifted_accum = final_accum + WGMMA::kNumAccum * local_idx;
#pragma unroll
for (int i = 0; i < WGMMA::kNumAccum / 4; ++ i) {
// NOTES: for unrolled `num_former_iters` cases, we expect the compiler to automatically make it a constant
bool predicate = kMustUseUniformedScaleB or i < num_former_iters;
shifted_accum[i * 4 + 0] += (predicate ? scale_0_0 : scale_0_1) * accum[i * 4 + 0];
shifted_accum[i * 4 + 1] += (predicate ? scale_0_0 : scale_0_1) * accum[i * 4 + 1];
shifted_accum[i * 4 + 2] += (predicate ? scale_1_0 : scale_1_1) * accum[i * 4 + 2];
shifted_accum[i * 4 + 3] += (predicate ? scale_1_0 : scale_1_1) * accum[i * 4 + 3];
}
for (uint32_t s = 0; s < kNumStages; ++ s) {
full_barriers[s]->wait((scheduler.current_iter * kNumIterations + k_iter) & 1);
empty_barrier_arrive(s);
}
}
}, num_former_iters);
} else {
// Launch MMAs
launch_k_iterations([&](int k_iter, auto type, auto num_former_iters_type) {
constexpr bool kHasDivisibleStages = std::is_same_v<decltype(type), DivisibleK>;
constexpr int kNumInnerStages = kHasDivisibleStages ? kNumStages : (SHAPE_K % kFullKOfAllStages) / BLOCK_K;
DG_STATIC_ASSERT(kNumInnerStages != 0, "Invalid number of inner stages");
// Wait unaligned cases
#pragma unroll
for (uint32_t s = kNumInnerStages; s < kNumStages; ++ s) {
full_barriers[s]->wait((scheduler.current_iter * kNumIterations + k_iter) & 1);
empty_barrier_arrive(s);
}
}, num_former_iters);
#pragma unroll
for (int s = 0; s < kNumInnerStages; ++ s) {
// Read B scales
float scale_b_0 = ld_shared(smem_scales_b + k_iter * kNumStages + s), scale_b_1;
// NOTES: even some blocks do not need to read the second row, but we still load one to align with other blocks
if constexpr (not kMustUseUniformedScaleB)
scale_b_1 = ld_shared(smem_scales_b + k_iter * kNumStages + s + SHAPE_K_SCALES);
// Wait TMA arrivals
full_barriers[s]->wait((scheduler.current_iter * kNumIterations + k_iter) & 1);
#pragma unroll
for (uint32_t local_idx = 0; local_idx < BLOCK_M / WAVE_BLOCK_M; ++ local_idx) {
auto m_offset = local_idx * WAVE_BLOCK_M;
// Read A scales
// NOTES: all shared memory read must be prior to `warpgroup_arrive` to avoid next scheduled block polluting the results
auto scale_a_0 = ld_shared(smem_scales_a[s] + r_0 + m_offset);
auto scale_a_1 = ld_shared(smem_scales_a[s] + r_1 + m_offset);
// Commit WGMMA instructions
#pragma unroll
for (int i = 0; i < WGMMA::kNumAccum; ++ i)
warpgroup_fence_operand(accum[i]);
warpgroup_arrive();
#pragma unroll
for (int k = 0; k < BLOCK_K / WGMMA::K; ++ k) {
auto desc_a = make_smem_desc(smem_a[s] + (math_wg_idx * WGMMA::M + m_offset) * BLOCK_K + k * WGMMA::K, 1);
auto desc_b = make_smem_desc(smem_b[s] + k * WGMMA::K, 1);
WGMMA::wgmma(desc_a, desc_b, accum, k);
}
warpgroup_commit_batch();
#pragma unroll
for (int i = 0; i < WGMMA::kNumAccum; ++ i)
warpgroup_fence_operand(accum[i]);
warpgroup_wait<0>();
// Notify barrier arrival at the last warpgroup wave
if (local_idx == BLOCK_M / WAVE_BLOCK_M - 1)
empty_barrier_arrive(s);
// Promote with scales
// NOTES: making it as predicates is very important for performance, comparing to two loops
float scale_0_0 = scale_a_0 * scale_b_0, scale_1_0 = scale_a_1 * scale_b_0;
float scale_0_1, scale_1_1;
if constexpr (not kMustUseUniformedScaleB)
scale_0_1 = scale_a_0 * scale_b_1, scale_1_1 = scale_a_1 * scale_b_1;
auto shifted_accum = final_accum + WGMMA::kNumAccum * local_idx;
#pragma unroll
for (int i = 0; i < WGMMA::kNumAccum / 4; ++ i) {
// NOTES: for unrolled `num_former_iters` cases, we expect the compiler to automatically make it a constant
bool predicate = kMustUseUniformedScaleB or i < num_former_iters;
shifted_accum[i * 4 + 0] += (predicate ? scale_0_0 : scale_0_1) * accum[i * 4 + 0];
shifted_accum[i * 4 + 1] += (predicate ? scale_0_0 : scale_0_1) * accum[i * 4 + 1];
shifted_accum[i * 4 + 2] += (predicate ? scale_1_0 : scale_1_1) * accum[i * 4 + 2];
shifted_accum[i * 4 + 3] += (predicate ? scale_1_0 : scale_1_1) * accum[i * 4 + 3];
}
}
}
// Wait unaligned cases
#pragma unroll
for (uint32_t s = kNumInnerStages; s < kNumStages; ++ s) {
full_barriers[s]->wait((scheduler.current_iter * kNumIterations + k_iter) & 1);
empty_barrier_arrive(s);
}
}, num_former_iters);
}
// TMA checks
constexpr uint32_t kNumElemBytes = sizeof(nv_bfloat16);

10
deep_gemm/include/deep_gemm/scheduler.cuh
View File

@ -48,6 +48,16 @@ struct Scheduler {
}
}
__device__ __forceinline__ bool is_valid_m(const uint32_t m_offset, const uint32_t& m_block_idx) const {
if constexpr (kGemmType == GemmType::Normal) {
return true;
} else if constexpr (kGemmType == GemmType::GroupedContiguous) {
return __ldg(grouped_layout + m_offset + m_block_idx * BLOCK_M) != -1;
} else if constexpr (kGemmType == GemmType::GroupedMasked) {
return m_offset + m_block_idx * BLOCK_M < __ldg(grouped_layout + curr_group_idx);
}
}
__device__ __forceinline__ bool is_tma_multicast_valid(const uint32_t& m_block_idx) const {
if (num_blocks_in_group == 1)
return false;

38
tests/test_core.py
View File

@ -53,8 +53,8 @@ def construct_contiguous_grouped(num_groups: int, expected_m_per_group: int, k:
m_aligned = get_m_alignment_for_contiguous_layout()
group_m_list = []
for i in range(num_groups):
group_m = m_aligned * random.randint(int(expected_m_per_group * 0.7) // m_aligned, int(expected_m_per_group * 1.3) // m_aligned)
m += group_m
group_m = random.randint(int(expected_m_per_group * 0.7), int(expected_m_per_group * 1.3))
m += ceil_div(group_m, m_aligned) * m_aligned
group_m_list.append(group_m)
x = torch.randn((m, k), device='cuda', dtype=torch.bfloat16)
y = torch.randn((num_groups, n, k), device='cuda', dtype=torch.bfloat16)
@ -64,10 +64,12 @@ def construct_contiguous_grouped(num_groups: int, expected_m_per_group: int, k:
start = 0
for i, group_m in enumerate(group_m_list):
end = start + group_m
m_indices[start:end] = i
ref_out[start:end] = x[start:end] @ y[i].t()
start = end
actual_end = start + group_m
aligned_end = start + ceil_div(group_m, m_aligned) * m_aligned
m_indices[start:actual_end] = i
m_indices[actual_end:aligned_end] = -1
ref_out[start:aligned_end] = x[start:aligned_end] @ y[i].t()
start = aligned_end
assert m % 4 == 0, f'TMA alignment error: {m}'
x_fp8 = per_token_cast_to_fp8(x)
@ -191,6 +193,8 @@ def test_m_grouped_gemm_contiguous() -> None:
# TODO: make a stronger test
m, x_fp8, y_fp8, m_indices, out, ref_out = construct_contiguous_grouped(num_groups, expected_m_per_group, k, n)
deep_gemm.m_grouped_gemm_fp8_fp8_bf16_nt_contiguous(x_fp8, y_fp8, out, m_indices)
out = torch.where((m_indices == -1).unsqueeze(1), torch.zeros_like(out), out)
ref_out = torch.where((m_indices == -1).unsqueeze(1), torch.zeros_like(ref_out), ref_out)
diff = calc_diff(out, ref_out)
assert diff < 0.001, f'{m=}, {k=}, {n=}, {diff:.5f}'
@ -202,24 +206,25 @@ def test_m_grouped_gemm_contiguous() -> None:
deep_gemm.m_grouped_gemm_fp8_fp8_bf16_nt_contiguous(x_fp8, y_fp8, out, m_indices)
t = bench_kineto(test_func, 'fp8_gemm', suppress_kineto_output=True)
sum_m = (m_indices != -1).sum().item()
print(f' > Performance ({num_groups=}, m={m:4}, n={n:4}, k={k:4}): {t * 1e6:4.0f} us | '
f'throughput: {2 * m * n * k / t / 1e12:4.0f} TFLOPS, '
f'{(m * k + num_groups * k * n + m * n * 2) / 1e9 / t:4.0f} GB/s')
f'throughput: {2 * sum_m * n * k / t / 1e12:4.0f} TFLOPS, '
f'{(sum_m * k + num_groups * k * n + sum_m * n * 2) / 1e9 / t:4.0f} GB/s')
print()
def test_m_grouped_gemm_masked() -> None:
print('Testing grouped masked GEMM:')
for num_groups, m in ((1, 1024), (2, 512), (4, 256)):
m = 4096
for num_groups, excepted_m in ((1, 1024), (2, 512), (4, 256)):
for k, n in ((7168, 4096), (2048, 7168), ):
# Test correctness
masked_m_candidates = list(filter(lambda candidate: candidate <= m, (64, 128, 192, 256, 320, 384)))
for i in range(10):
x_fp8, y_fp8, out, ref_out = construct_masked_grouped(num_groups, m, k, n)
masked_m = torch.empty((num_groups, ), device='cuda', dtype=torch.int)
for j in range(num_groups):
masked_m[j] = random.choice(masked_m_candidates)
masked_m[j] = random.randint(int(excepted_m * 0.7), int(excepted_m * 1.3))
expected_m = min(int(masked_m.float().mean()) + 1, m)
deep_gemm.m_grouped_gemm_fp8_fp8_bf16_nt_masked(x_fp8, y_fp8, out, masked_m, expected_m)
for j in range(num_groups):
@ -228,17 +233,20 @@ def test_m_grouped_gemm_masked() -> None:
# Construct new tensors only once to avoid L2 cache acceleration (creating them puts them in L2)
x_fp8, y_fp8, out, ref_out = construct_masked_grouped(num_groups, m, k, n)
masked_m = torch.ones((num_groups, ), device='cuda', dtype=torch.int) * m
for j in range(num_groups):
masked_m[j] = random.randint(int(excepted_m * 0.7), int(excepted_m * 1.3))
expected_m = min(int(masked_m.float().mean()) + 1, m)
sum_m = masked_m.sum().item()
# noinspection PyShadowingNames
def test_func():
deep_gemm.m_grouped_gemm_fp8_fp8_bf16_nt_masked(x_fp8, y_fp8, out, masked_m, m)
deep_gemm.m_grouped_gemm_fp8_fp8_bf16_nt_masked(x_fp8, y_fp8, out, masked_m, expected_m)
# Test performance with fixed shapes
t = bench_kineto(test_func, 'fp8_gemm', suppress_kineto_output=True)
print(f' > Performance ({num_groups=}, m_per_group={m:4}, n={n:4}, k={k:4}): {t * 1e6:4.0f} us | '
f'throughput: {2 * num_groups * m * n * k / t / 1e12:4.0f} TFLOPS, '
f'{(num_groups * (m * k + k * n + m * n * 2)) / 1e9 / t:4.0f} GB/s')
f'throughput: {2 * sum_m * n * k / t / 1e12:4.0f} TFLOPS, '
f'{(sum_m * k + num_groups * k * n + sum_m * n * 2) / 1e9 / t:4.0f} GB/s')
print()