Extend support for ak1 / bk1 WMMA (#3073)

* Extend AK1 / BK1 support:

 - Add support for AK1 != BK1
 - Add support for AK1, BK1 > 8
 - Introduce KInner template parameter for pipelines when loading multiple tiles with one instruction

* fix clang format
This commit is contained in:
Enrico Degregori
2025-11-11 16:38:15 +01:00
committed by GitHub
parent 9f33b7cfd3
commit 1c544abf57
24 changed files with 621 additions and 339 deletions

View File

@@ -95,6 +95,7 @@ struct wmma_type<WmmaInstr::wmma_f32_16x16x16_f16,
static constexpr index_t m_per_wmma = 16;
static constexpr index_t n_per_wmma = 16;
static constexpr index_t k_per_wmma = 16;
static constexpr index_t k_per_blk = 8;
static constexpr index_t src_a_data_size = 2;
static constexpr index_t src_b_data_size = 2;
static constexpr index_t acc_data_size = 4;
@@ -136,6 +137,7 @@ struct wmma_type<WmmaInstr::wmma_f32_16x16x16_bf16,
static constexpr index_t m_per_wmma = 16;
static constexpr index_t n_per_wmma = 16;
static constexpr index_t k_per_wmma = 16;
static constexpr index_t k_per_blk = 8;
static constexpr index_t src_a_data_size = 2;
static constexpr index_t src_b_data_size = 2;
static constexpr index_t acc_data_size = 4;
@@ -173,6 +175,7 @@ struct wmma_type<WmmaInstr::wmma_f16_16x16x16_f16,
static constexpr index_t m_per_wmma = 16;
static constexpr index_t n_per_wmma = 16;
static constexpr index_t k_per_wmma = 16;
static constexpr index_t k_per_blk = 8;
static constexpr index_t src_a_data_size = 2;
static constexpr index_t src_b_data_size = 2;
static constexpr index_t acc_data_size = 2;
@@ -209,6 +212,7 @@ struct wmma_type<WmmaInstr::wmma_bf16_16x16x16_bf16,
static constexpr index_t m_per_wmma = 16;
static constexpr index_t n_per_wmma = 16;
static constexpr index_t k_per_wmma = 16;
static constexpr index_t k_per_blk = 8;
static constexpr index_t src_a_data_size = 2;
static constexpr index_t src_b_data_size = 2;
static constexpr index_t acc_data_size = 2;
@@ -251,6 +255,7 @@ struct wmma_type<WmmaInstr::wmma_i32_16x16x16_iu8,
static constexpr index_t m_per_wmma = 16;
static constexpr index_t n_per_wmma = 16;
static constexpr index_t k_per_wmma = 16;
static constexpr index_t k_per_blk = 8;
static constexpr index_t src_a_data_size = 2;
static constexpr index_t src_b_data_size = 2;
static constexpr index_t acc_data_size = 4;
@@ -301,6 +306,7 @@ struct wmma_type<WmmaInstr::wmma_f32_16x16x16_f16_gfx12,
static constexpr index_t m_per_wmma = 16;
static constexpr index_t n_per_wmma = 16;
static constexpr index_t k_per_wmma = 16;
static constexpr index_t k_per_blk = 8;
// static constexpr index_t src_a_data_size = 2;
// static constexpr index_t src_b_data_size = 2;
// static constexpr index_t acc_data_size = 4;
@@ -339,6 +345,7 @@ struct wmma_type<WmmaInstr::wmma_f32_16x16x16_bf16_gfx12,
static constexpr index_t m_per_wmma = 16;
static constexpr index_t n_per_wmma = 16;
static constexpr index_t k_per_wmma = 16;
static constexpr index_t k_per_blk = 8;
// static constexpr index_t src_a_data_size = 2;
// static constexpr index_t src_b_data_size = 2;
static constexpr index_t acc_data_size = 4;
@@ -372,6 +379,7 @@ struct wmma_type<WmmaInstr::wmma_i32_16x16x16_iu8_gfx12,
static constexpr index_t m_per_wmma = 16;
static constexpr index_t n_per_wmma = 16;
static constexpr index_t k_per_wmma = 16;
static constexpr index_t k_per_blk = 8;
// static constexpr index_t src_a_data_size = 2;
// static constexpr index_t src_b_data_size = 2;
static constexpr index_t acc_data_size = 4;
@@ -413,6 +421,7 @@ struct wmma_type<WmmaInstr::wmma_f32_16x16x16_f8f8_gfx12,
static constexpr index_t m_per_wmma = 16;
static constexpr index_t n_per_wmma = 16;
static constexpr index_t k_per_wmma = 16;
static constexpr index_t k_per_blk = 8;
static constexpr index_t acc_data_size = 4;
static constexpr index_t acc_pack_number = 1;
static constexpr index_t num_thread_per_subgroups = n_per_wmma;
@@ -448,6 +457,7 @@ struct wmma_type<WmmaInstr::wmma_f32_16x16x16_f8bf8_gfx12,
static constexpr index_t m_per_wmma = 16;
static constexpr index_t n_per_wmma = 16;
static constexpr index_t k_per_wmma = 16;
static constexpr index_t k_per_blk = 8;
static constexpr index_t acc_data_size = 4;
static constexpr index_t acc_pack_number = 1;
static constexpr index_t num_thread_per_subgroups = n_per_wmma;
@@ -483,6 +493,7 @@ struct wmma_type<WmmaInstr::wmma_f32_16x16x16_bf8f8_gfx12,
static constexpr index_t m_per_wmma = 16;
static constexpr index_t n_per_wmma = 16;
static constexpr index_t k_per_wmma = 16;
static constexpr index_t k_per_blk = 8;
static constexpr index_t acc_data_size = 4;
static constexpr index_t acc_pack_number = 1;
static constexpr index_t num_thread_per_subgroups = n_per_wmma;
@@ -518,6 +529,7 @@ struct wmma_type<WmmaInstr::wmma_f32_16x16x16_bf8bf8_gfx12,
static constexpr index_t m_per_wmma = 16;
static constexpr index_t n_per_wmma = 16;
static constexpr index_t k_per_wmma = 16;
static constexpr index_t k_per_blk = 8;
static constexpr index_t acc_data_size = 4;
static constexpr index_t acc_pack_number = 1;
static constexpr index_t num_thread_per_subgroups = n_per_wmma;
@@ -768,6 +780,8 @@ struct WmmaGemm
__device__ static constexpr index_t GetWaveSize() { return wmma_instr.wave_size; }
__device__ static constexpr index_t GetKPerWaveBlk() { return wmma_instr.k_per_blk; }
template <class FloatA, class FloatB, class FloatC>
__device__ void Run(const FloatA& p_a_wave, const FloatB& p_b_wave, FloatC& p_c_thread) const
{