flash: Fix rowsum and write fake exp

GEMM part is disabled for faster debugging, the kernel reads the result
of A*B directly from input binary.
This commit is contained in:
Hansung Kim
2024-08-15 16:32:21 -07:00
parent 53dfc690b9
commit e809d25305

View File

@@ -2,6 +2,7 @@
#include <vx_intrinsics.h>
#include <vx_print.h>
#include <vx_spawn.h>
#include <float.h>
#include "common.h"
#include "sgemm_impl.hpp"
#include "include/gemmini.h"
@@ -13,7 +14,7 @@ using float_type = float16_t;
#define B_ROW BM
#define B_COL BN
inline void thread_block_flashattn(float *S, float *gmem,
inline void thread_block_flashattn(float *S,
const uint32_t tid_in_threadblock,
const uint32_t threads_per_threadblock,
const uint32_t threadblock_id_in_cluster,
@@ -37,29 +38,121 @@ inline void thread_block_flashattn(float *S, float *gmem,
// asm volatile("fmv.s %0, f22" : "=f"(ft[6]));
// asm volatile("fmv.s %0, f23" : "=f"(ft[7]));
// row-max
//
// one warp handles one row in tile; iterate enough times to cover all the
// rows
volatile float *gmem_tmp0 = reinterpret_cast<volatile float *>(0xd0000000UL);
volatile float *gmem_tmp1 = reinterpret_cast<volatile float *>(0xe0000000UL);
volatile float *gmem_tmp2 = reinterpret_cast<volatile float *>(0xf0000000UL);
for (int warp_offset = 0; warp_offset < B_ROW;
warp_offset += warps_in_threadblock) {
const uint32_t row = warp_offset + warp_id;
const uint32_t first_thread_offset = B_COL * row;
uint32_t thread_offset = first_thread_offset + tid_in_warp;
float curr_max = S[first_thread_offset];
constexpr uint32_t load_iter = B_COL / NUM_THREADS;
// rowmax
//
// two-level tree reduction: reduce each row into NUM_THREADS intermediate
// maxes, then reduce it to one global max
// one warp handles one row in tile
// #define DUMB_ROWMAX
#ifdef DUMB_ROWMAX
if (tid_in_warp == 0) {
float max = S[first_thread_offset];
#pragma GCC unroll
for (int iter = 0; iter < load_iter; iter++) {
for (int i = 0; i < B_COL; i++) {
asm volatile("fmax.s %0, %1, %2"
: "=f"(max)
: "f"(max), "f"(S[first_thread_offset + i]));
}
sharedmem_row_max_sum[row] = max;
gmem_tmp0[row] = max;
}
#else
static_assert((B_ROW % NUM_THREADS) == 0,
"B_ROW must be a multiple of NUM_THREADS");
constexpr uint32_t per_row_iter = B_COL / NUM_THREADS;
uint32_t thread_offset = first_thread_offset + tid_in_warp;
float per_thread_max = FLT_MIN;
#pragma GCC unroll
for (int i = 0; i < per_row_iter; i++) {
const float next = S[thread_offset];
asm volatile("fmax.s %0, %1, %2"
: "=f"(curr_max)
: "f"(curr_max), "f"(S[thread_offset]));
: "=f"(per_thread_max)
: "f"(per_thread_max), "f"(next));
thread_offset += NUM_THREADS;
}
// get max value across the same-warp threads using smem
// NOTE: be careful with out-of-bounds
// stage per-thread max value in smem
// FIXME: we could warp_id instead of row here, but we need another barrier
// at the end of the loop iteration to prevent write-after-read hazard
// FIXME: threadblock_id needs to be in here too
float *warp_smem = sharedmem_scratchpad + (row * NUM_THREADS);
warp_smem[tid_in_warp] = curr_max;
warp_smem[tid_in_warp] = per_thread_max;
// sync writes to warp_smem
threadblock_barrier(threadblock_id_in_cluster,
warps_per_threadblock_per_core);
// elect 0-th thread to reduce all other thread's values in the warp
if (tid_in_warp == 0) {
for (int iter = 1; iter < NUM_THREADS; iter++) {
float other = warp_smem[iter];
asm volatile("fmax.s %0, %1, %2"
: "=f"(per_thread_max)
: "f"(per_thread_max), "f"(other));
}
sharedmem_row_max_sum[row] = per_thread_max;
gmem_tmp0[row] = per_thread_max;
}
#endif
// FIXME: unnecessary?
threadblock_barrier(threadblock_id_in_cluster,
warps_per_threadblock_per_core);
// exponential
//
// B_ROW / (B_ROW * B_COL / (exp_elem * threads_per_threadblock))
// const uint32_t row_stride =
// (exp_elem_per_thread * threads_per_threadblock) / B_COL;
thread_offset = first_thread_offset + tid_in_warp;
// broadcast rowmax to all threads in the warp
const float row_max = sharedmem_row_max_sum[row];
#pragma GCC unroll
for (int i = 0; i < per_row_iter; i++) {
float val = S[thread_offset];
// FIXME: placeholder for proper exp
val = val;
// update S in-place to P
// S[thread_offset] = val;
gmem_tmp1[thread_offset] = val;
gmem_tmp2[thread_offset] = val - row_max;
thread_offset += NUM_THREADS;
}
threadblock_barrier(threadblock_id_in_cluster,
warps_per_threadblock_per_core);
// rowsum
//
// two-level tree reduction, similar to rowmax
#if 0
float per_thread_sum = 0.0f;
#pragma GCC unroll
for (int i = 0; i < per_row_iter; i++) {
per_thread_sum += S[thread_offset];
thread_offset += NUM_THREADS;
}
// stage per-thread sum value in smem
// FIXME: threadblock_id needs to be in here too
warp_smem = sharedmem_scratchpad + (row * NUM_THREADS);
warp_smem[tid_in_warp] = per_thread_sum;
// sync writes to warp_smem
threadblock_barrier(threadblock_id_in_cluster,
@@ -69,21 +162,16 @@ inline void thread_block_flashattn(float *S, float *gmem,
if (tid_in_warp == 0) {
for (int iter = 1; iter < NUM_THREADS; iter++) {
float other = warp_smem[iter];
asm volatile("fmax.s %0, %1, %2"
: "=f"(curr_max)
: "f"(curr_max), "f"(other));
per_thread_sum += other;
}
sharedmem_row_max_sum[row] = curr_max;
sharedmem_row_max_sum[row] = per_thread_sum;
gmem_tmp2[row] = per_thread_sum;
}
}
#endif
// exponential
//
// FIXME: placeholder for proper exp
constexpr uint32_t exp_elem_per_thread = 1;
// B_ROW / (B_ROW * B_COL / (exp_elem * threads_per_threadblock))
const uint32_t row_stride =
(exp_elem_per_thread * threads_per_threadblock) / B_COL;
threadblock_barrier(threadblock_id_in_cluster,
warps_per_threadblock_per_core);
}
asm volatile("thread_block_flashattn_finish_%=:" ::);
}
@@ -135,17 +223,18 @@ void kernel_body(int task_id, kernel_arg_t *__UNIFORM__ arg) {
reinterpret_cast<uint8_t *>(SMEM_ADDR_END) - sharedmem_row_max_sum_size;
// sharedmem "scratchpad" area to put temporary data, e.g. for tree reduction
// in rowsum
// FIXME: size is arbitrary, and out-of bounds is not checked
constexpr uint32_t sharedmem_scratchpad_size = 0x1000;
// NOTE: out-of bounds is not checked
constexpr uint32_t sharedmem_scratchpad_size =
sizeof(float) * B_ROW * NUM_THREADS * 2 /*arbitrary slack*/;
uint8_t *sharedmem_scratchpad =
sharedmem_row_max_sum - sharedmem_scratchpad_size;
thread_block_gemm<float_type, /*write_to_gmem=*/true>(
(const float_type *)arg->addr_a, (const float_type *)arg->addr_b,
(float *)smem_S /*write result to SMEM */, arg->dim_m, arg->dim_n,
arg->dim_k, tid_in_threadblock, threads_per_threadblock,
threadblocks_per_cluster, threadblock_id_in_cluster,
sharedmem_per_threadblock);
// thread_block_gemm<float_type, /*write_to_gmem=*/true>(
// (const float_type *)arg->addr_a, (const float_type *)arg->addr_b,
// (float *)smem_S /*write result to SMEM */, arg->dim_m, arg->dim_n,
// arg->dim_k, tid_in_threadblock, threads_per_threadblock,
// threadblocks_per_cluster, threadblock_id_in_cluster,
// sharedmem_per_threadblock);
// protect writes of GEMM results before softmax
const uint32_t warps_per_threadblock_per_core =
@@ -153,10 +242,10 @@ void kernel_body(int task_id, kernel_arg_t *__UNIFORM__ arg) {
threadblock_barrier(threadblock_id_in_cluster,
warps_per_threadblock_per_core);
thread_block_flashattn(
(float *)smem_S, (float *)arg->addr_c, tid_in_threadblock,
threads_per_threadblock, threadblock_id_in_cluster,
(float *)sharedmem_scratchpad_size, (float *)sharedmem_row_max_sum);
thread_block_flashattn((float *)arg->addr_a /* smem_S, */, tid_in_threadblock,
threads_per_threadblock, threadblock_id_in_cluster,
(float *)sharedmem_scratchpad,
(float *)sharedmem_row_max_sum);
}
int main() {