wu-arch/kernels
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Switch header configs to flash

Browse Source
This commit is contained in:
Hansung Kim
2024-11-08 21:56:26 -08:00
parent 4e087a8aab
commit 1e3d476e70
3 changed files with 9 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
kernel/include/gemmini_mmio.h
View File

@@ -9,9 +9,9 @@
// #define SMEM_SIZE 0x4000
// 64KB
// #define SMEM_SIZE 0x10000
// 128KB
// 128KB (FP16 GEMM)
// #define SMEM_SIZE 0x20000
// 256KB
// 256KB (FlashAttention)
#define SMEM_SIZE 0x40000
#define SMEM_MASK (SMEM_SIZE - 1)

6
tests/regression/flash_attention/flash_impl.hpp
View File

@@ -11,8 +11,10 @@
#define ROW_REMAINDER_LOGIC
constexpr uint32_t ROWMAX_SETS = 3;
constexpr bool WARP_SPECIALIZED = true;
constexpr bool TENSOR_CORE = true;
// constexpr bool WARP_SPECIALIZED = true;
// constexpr bool TENSOR_CORE = true;
constexpr bool WARP_SPECIALIZED = false;
constexpr bool TENSOR_CORE = false;
// temporary safety stop for wrong configs
static_assert(NUM_CORES == 4);

6
tests/regression/sgemm_tcore/sgemm_impl.hpp
View File

@@ -6,7 +6,7 @@
#include "include/gemmini.h"
#include "gemmini_mmio.h"
#define FP_SIZE 16
#define FP_SIZE 32
// "fake" fp16 type that only has the correct data width.
using float16_t = uint16_t;
@@ -19,7 +19,7 @@ using float_type = float16_t;
// Generate kernel for the Hopper-style SMEM-decoupled tensor core. This uses
// asynchronous HGMMA and HGMMA_WAIT instructions.
#define TENSOR_HOPPER 1
#define TENSOR_HOPPER 0
// Constraints on parameters:
// * Memory:
@@ -110,7 +110,7 @@ static_assert(WMITER * WNITER * TCM * TCN * NUM_WARPS * CORES_PER_CLUSTER ==
// result matrix will be stored in a swizzled form in the global memory.
#define WMMA_STORE_FAST 1
#define GEMMINI_DMA 0
#define GEMMINI_DMA 1
#define GEMMINI_DMA_FAST 1
#if SMEM_SIZE == 0x4000
#define SMEM_ADDR_Q0 ((float * const) 0xff000000)