441 lines
16 KiB
C++
441 lines
16 KiB
C++
#include <stdint.h>
|
|
#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"
|
|
#include "gemmini_mmio.h"
|
|
|
|
#define B_ROW BM
|
|
#define B_COL BN
|
|
// FIXME
|
|
#define HEADDIM B_COL
|
|
|
|
inline void thread_block_init_sharedmem(const uint32_t tid_in_threadblock,
|
|
const uint32_t threads_per_threadblock,
|
|
float *smem_O,
|
|
float *smem_rowmax,
|
|
float *smem_rowsum) {
|
|
const uint32_t tid_in_warp = tid_in_threadblock % NUM_THREADS;
|
|
const uint32_t warp_id = tid_in_threadblock / NUM_THREADS;
|
|
const uint32_t warps_in_threadblock = threads_per_threadblock / NUM_THREADS;
|
|
|
|
static_assert((B_ROW % NUM_THREADS) == 0,
|
|
"B_ROW must be a multiple of NUM_THREADS");
|
|
// FIXME: this shouldn't be necessary
|
|
static_assert(B_ROW < (NUM_THREADS * CORES_PER_CLUSTER * NUM_WARPS),
|
|
"not enough warps to initialize rowmax/rowsum");
|
|
|
|
constexpr uint32_t num_warps = B_ROW / NUM_THREADS;
|
|
if (warp_id < num_warps) {
|
|
uint32_t offset = NUM_THREADS * warp_id + tid_in_warp;
|
|
// mi, mi~, minew
|
|
smem_rowmax[offset] = FLT_MIN;
|
|
smem_rowmax[offset + B_ROW] = FLT_MIN;
|
|
smem_rowmax[offset + 2 * B_ROW] = FLT_MIN;
|
|
smem_rowsum[offset] = 0.0f;
|
|
}
|
|
|
|
// FIXME: dedup this pattern
|
|
for (int warp_offset = 0; warp_offset < B_COL;
|
|
warp_offset += warps_in_threadblock) {
|
|
// each warp clears out a row of smem_O
|
|
const uint32_t row = warp_offset + warp_id;
|
|
uint32_t thread_offset = HEADDIM * row + tid_in_warp;
|
|
constexpr uint32_t per_row_iter = HEADDIM / NUM_THREADS;
|
|
#pragma GCC unroll
|
|
for (int i = 0; i < per_row_iter; i++) {
|
|
smem_O[thread_offset] = 0.0f;
|
|
thread_offset += NUM_THREADS;
|
|
}
|
|
}
|
|
}
|
|
|
|
inline void thread_block_online_softmax(
|
|
const float *smem_S, float *smem_O, float *smem_P,
|
|
const uint32_t tid_in_threadblock, const uint32_t threads_per_threadblock,
|
|
const uint32_t threadblocks_per_cluster,
|
|
const uint32_t threadblock_id_in_cluster, float *smem_scratchpad,
|
|
float *smem_rowmax, float *smem_rowsum) {
|
|
asm volatile("thread_block_flashattn_start_%=:" ::);
|
|
|
|
const uint32_t tid_in_warp = tid_in_threadblock % NUM_THREADS;
|
|
const uint32_t warp_id = tid_in_threadblock / NUM_THREADS;
|
|
const uint32_t warps_in_threadblock = threads_per_threadblock / NUM_THREADS;
|
|
const uint32_t warps_per_threadblock_per_core =
|
|
NUM_WARPS / threadblocks_per_cluster;
|
|
|
|
// float ft[8];
|
|
// asm volatile("fmv.s %0, f16" : "=f"(ft[0]));
|
|
// asm volatile("fmv.s %0, f17" : "=f"(ft[1]));
|
|
// asm volatile("fmv.s %0, f18" : "=f"(ft[2]));
|
|
// asm volatile("fmv.s %0, f19" : "=f"(ft[3]));
|
|
// asm volatile("fmv.s %0, f20" : "=f"(ft[4]));
|
|
// asm volatile("fmv.s %0, f21" : "=f"(ft[5]));
|
|
// asm volatile("fmv.s %0, f22" : "=f"(ft[6]));
|
|
// asm volatile("fmv.s %0, f23" : "=f"(ft[7]));
|
|
|
|
volatile float *gmem_tmp0 = reinterpret_cast<volatile float *>(0xd0000000UL);
|
|
volatile float *gmem_tmp1 = reinterpret_cast<volatile float *>(0xe0000000UL);
|
|
|
|
float *smem_rowmax_prev = smem_rowmax;
|
|
float *smem_rowmax_new = smem_rowmax + B_ROW;
|
|
float *smem_rowmax_this = smem_rowmax + 2 * B_ROW;
|
|
|
|
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;
|
|
|
|
// 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 i = 0; i < B_COL; i++) {
|
|
asm volatile("fmax.s %0, %1, %2"
|
|
: "=f"(max)
|
|
: "f"(max), "f"(S[first_thread_offset + i]));
|
|
}
|
|
smem_rowmax[row] = max;
|
|
}
|
|
|
|
#else
|
|
static_assert((B_COL % NUM_THREADS) == 0,
|
|
"B_COL 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 = smem_S[thread_offset];
|
|
asm volatile("fmax.s %0, %1, %2"
|
|
: "=f"(per_thread_max)
|
|
: "f"(per_thread_max), "f"(next));
|
|
thread_offset += NUM_THREADS;
|
|
}
|
|
// stage per-thread max value in smem
|
|
// FIXME: threadblock_id needs to be in here too
|
|
float *warp_smem = smem_scratchpad + (warp_id * NUM_THREADS);
|
|
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) {
|
|
float rowmax = per_thread_max;
|
|
for (int iter = 1; iter < NUM_THREADS; iter++) {
|
|
float other = warp_smem[iter];
|
|
asm volatile("fmax.s %0, %1, %2"
|
|
: "=f"(rowmax)
|
|
: "f"(rowmax), "f"(other));
|
|
}
|
|
smem_rowmax_this[row] = rowmax;
|
|
|
|
// update previous rowmax
|
|
// i.e. mi_new = max(mi, mij)
|
|
float prev_rowmax = smem_rowmax_prev[row];
|
|
asm volatile("fmax.s %0, %1, %2"
|
|
: "=f"(rowmax)
|
|
: "f"(rowmax), "f"(prev_rowmax));
|
|
smem_rowmax_new[row] = rowmax;
|
|
}
|
|
#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;
|
|
|
|
// broadcast rowmax to all threads in the warp
|
|
const float rowmax_new = smem_rowmax_new[row];
|
|
|
|
// each thread computes two fp32 elements, downconverts it to fp16, then
|
|
// packs them into one fp32
|
|
constexpr uint32_t elem_per_thread = 1;
|
|
static_assert((B_COL % (elem_per_thread * NUM_THREADS)) == 0,
|
|
"B_COL condition not met for P compute");
|
|
|
|
thread_offset = first_thread_offset + (elem_per_thread * tid_in_warp);
|
|
constexpr uint32_t exp_per_row_iter =
|
|
B_COL / (elem_per_thread * NUM_THREADS);
|
|
#pragma GCC unroll
|
|
for (int i = 0; i < exp_per_row_iter; i++) {
|
|
float f0 = smem_S[thread_offset];
|
|
|
|
// check Q*K result
|
|
gmem_tmp0[thread_offset] = f0;;
|
|
|
|
// FIXME: placeholder for proper exp
|
|
f0 -= rowmax_new;
|
|
|
|
// Store S transposed to the shared memory
|
|
|
|
smem_P[thread_offset] = f0;
|
|
gmem_tmp1[thread_offset] = f0;
|
|
|
|
thread_offset += NUM_THREADS;
|
|
}
|
|
|
|
threadblock_barrier(threadblock_id_in_cluster,
|
|
warps_per_threadblock_per_core);
|
|
|
|
// rowsum
|
|
//
|
|
// two-level tree reduction, similar to rowmax
|
|
|
|
thread_offset = first_thread_offset + tid_in_warp;
|
|
float per_thread_sum = 0.0f;
|
|
#pragma GCC unroll
|
|
for (int i = 0; i < per_row_iter; i++) {
|
|
per_thread_sum += smem_P[thread_offset];
|
|
thread_offset += NUM_THREADS;
|
|
}
|
|
// stage per-thread sum value in smem
|
|
// FIXME: threadblock_id needs to be in here too
|
|
warp_smem = smem_scratchpad + (warp_id * NUM_THREADS);
|
|
warp_smem[tid_in_warp] = per_thread_sum;
|
|
|
|
// sync writes to warp_smem
|
|
threadblock_barrier(threadblock_id_in_cluster,
|
|
warps_per_threadblock_per_core);
|
|
|
|
// 0-th thread collects all other thread's values in the warp
|
|
if (tid_in_warp == 0) {
|
|
float rowsum = per_thread_sum;
|
|
for (int iter = 1; iter < NUM_THREADS; iter++) {
|
|
float other = warp_smem[iter];
|
|
rowsum += other;
|
|
}
|
|
|
|
const float mi_prev = smem_rowmax_prev[row];
|
|
const float mi_this = smem_rowmax_this[row];
|
|
const float exp = mi_prev - mi_this;
|
|
|
|
// update rowsum
|
|
const float rowsum_prev = smem_rowsum[row];
|
|
// FIXME: placeholder for exponential
|
|
float rowsum_new = exp * rowsum_prev + rowsum;
|
|
smem_rowsum[row] = rowsum_new;
|
|
}
|
|
|
|
threadblock_barrier(threadblock_id_in_cluster,
|
|
warps_per_threadblock_per_core);
|
|
|
|
// Oi rescale
|
|
//
|
|
thread_offset = first_thread_offset + tid_in_warp;
|
|
#pragma GCC unroll
|
|
for (int i = 0; i < per_row_iter; i++) {
|
|
float fval = smem_O[thread_offset];
|
|
|
|
const float mi_prev = smem_rowmax_prev[row];
|
|
const float mi_new = smem_rowmax_new[row];
|
|
const float exp = mi_prev - mi_new;
|
|
|
|
// FIXME: placeholder for proper exp
|
|
fval *= exp;
|
|
|
|
// update Oi in-place
|
|
smem_O[thread_offset] = fval;
|
|
|
|
thread_offset += NUM_THREADS;
|
|
}
|
|
|
|
threadblock_barrier(threadblock_id_in_cluster,
|
|
warps_per_threadblock_per_core);
|
|
}
|
|
|
|
asm volatile("thread_block_flashattn_finish_%=:" ::);
|
|
}
|
|
|
|
void kernel_body(int task_id, kernel_arg_t *__UNIFORM__ arg) {
|
|
// @perf: All threads are running these compute whose result is mostly same
|
|
// across the threadblock
|
|
|
|
#ifdef RADIANCE
|
|
constexpr uint32_t cores_per_cluster = CORES_PER_CLUSTER;
|
|
#else
|
|
constexpr uint32_t cores_per_cluster = 1;
|
|
#endif
|
|
|
|
// FIXME: headdim not considered
|
|
uint32_t threads_per_threadblock = (B_ROW * B_COL) / (ELEM_PER_THREAD);
|
|
const uint32_t hw_threads_per_cluster =
|
|
cores_per_cluster * vx_num_threads() * vx_num_warps();
|
|
// cap maximum threadblock size to # of HW threads in cluster, to prevent
|
|
// multiple "wave" invocations which slows down the kernel
|
|
if (threads_per_threadblock > hw_threads_per_cluster) {
|
|
threads_per_threadblock = hw_threads_per_cluster;
|
|
}
|
|
const uint32_t threadblocks_per_cluster =
|
|
hw_threads_per_cluster / threads_per_threadblock;
|
|
|
|
const int threadblock_id = task_id / threads_per_threadblock;
|
|
const int threadblock_id_in_cluster =
|
|
threadblock_id % threadblocks_per_cluster;
|
|
const int tid_in_threadblock = task_id % threads_per_threadblock;
|
|
|
|
const uint32_t dim_seqlen = arg->dim_seqlen;
|
|
const uint32_t dim_headdim = arg->dim_headdim;
|
|
|
|
// "static" shared memory allocation. This would determine maximum
|
|
// threadblock occupancy in a cluster
|
|
constexpr uint32_t smem_Q_size = B_ROW * HEADDIM;
|
|
constexpr uint32_t smem_QK_size = B_ROW * B_COL;
|
|
constexpr uint32_t smem_V_size = B_COL * HEADDIM;
|
|
constexpr uint32_t smem_O_size = B_COL * HEADDIM;
|
|
uint8_t *smem_per_threadblock = reinterpret_cast<uint8_t *>(
|
|
DEV_SMEM_START_ADDR +
|
|
sizeof(float_type) *
|
|
(smem_QK_size + smem_V_size + smem_O_size) *
|
|
threadblock_id_in_cluster);
|
|
|
|
float *smem_Q = reinterpret_cast<float *>(smem_per_threadblock);
|
|
float *smem_K = smem_Q + smem_Q_size;
|
|
// in-place multiplication of QK into Q
|
|
float *smem_S = reinterpret_cast<float *>(smem_per_threadblock);
|
|
float *smem_P = smem_S; // in-place update from S to P
|
|
float *smem_V =
|
|
reinterpret_cast<float *>(smem_per_threadblock) + smem_QK_size;
|
|
float *smem_O = reinterpret_cast<float *>(smem_per_threadblock) +
|
|
smem_QK_size + smem_V_size;
|
|
|
|
// allocate rowmax/rowsum storage at the end of the sharedmem address space
|
|
constexpr uint32_t smem_rowmax_size = B_ROW * 3 /* mi, mi~, minew */;
|
|
constexpr uint32_t smem_rowsum_size = B_ROW;
|
|
float *smem_rowmax =
|
|
reinterpret_cast<float *>(SMEM_ADDR_END) - smem_rowmax_size;
|
|
float *smem_rowsum = smem_rowmax - smem_rowsum_size;
|
|
|
|
// sharedmem "scratchpad" area to put temporary data, e.g. for tree reduction
|
|
// in rowsum
|
|
// NOTE: out-of bounds is not checked
|
|
constexpr uint32_t smem_scratchpad_size =
|
|
B_ROW * NUM_THREADS * 2 /*arbitrary slack*/;
|
|
float *smem_scratchpad = smem_rowmax - smem_scratchpad_size;
|
|
|
|
const uint32_t warps_per_threadblock_per_core =
|
|
NUM_WARPS / threadblocks_per_cluster;
|
|
|
|
// initialize rowmax/rowsum values in sharedmem
|
|
thread_block_init_sharedmem(tid_in_threadblock, threads_per_threadblock,
|
|
smem_O, smem_rowmax, smem_rowsum);
|
|
|
|
const float *gmem_Q = reinterpret_cast<float *>(arg->addr_q);
|
|
const float *gmem_K = reinterpret_cast<float *>(arg->addr_k);
|
|
const float *gmem_V = reinterpret_cast<float *>(arg->addr_v);
|
|
float *gmem_O = reinterpret_cast<float *>(arg->addr_o);
|
|
float *gmem_tmp0 = reinterpret_cast<float *>(0xd0000000UL);
|
|
|
|
// #define SKIP_GEMM
|
|
#ifndef SKIP_GEMM
|
|
#if 0
|
|
thread_block_gemm<float_type, /*write_to_gmem=*/true>(
|
|
(const float_type *)arg->addr_q, (const float_type *)arg->addr_k,
|
|
(float *)smem_S /*write result to SMEM */, B_ROW, B_COL,
|
|
HEADDIM, tid_in_threadblock, threads_per_threadblock,
|
|
threadblocks_per_cluster, threadblock_id_in_cluster,
|
|
smem_per_threadblock);
|
|
|
|
#else
|
|
|
|
// clear out accumulators
|
|
initialize_accum_regs<0>();
|
|
initialize_accum_regs<1>();
|
|
|
|
// load Q
|
|
static_assert(B_ROW == B_COL, "currently only supports square tiles");
|
|
load_tile_to_smem<float, MemLayout::MN_major, MemLayout::MN_major, B_ROW,
|
|
HEADDIM>(B_ROW, 0, 0, gmem_Q, smem_Q, tid_in_threadblock);
|
|
// load K
|
|
load_tile_to_smem<float, MemLayout::MN_major, MemLayout::MN_major, B_COL,
|
|
HEADDIM>(B_COL, 0, 0, gmem_K, smem_K, tid_in_threadblock);
|
|
|
|
// GMEM->SMEM and compute barrier
|
|
threadblock_barrier(threadblock_id_in_cluster,
|
|
warps_per_threadblock_per_core);
|
|
|
|
thread_block_gemm_single_tile<float, MemLayout::MN_major, MemLayout::MN_major,
|
|
/*write_to_smem=*/true>(
|
|
smem_Q, smem_K, smem_S, tid_in_threadblock, threads_per_threadblock,
|
|
threadblocks_per_cluster, threadblock_id_in_cluster);
|
|
#endif
|
|
|
|
// protect GEMM result writes (smem_S) before softmax
|
|
threadblock_barrier(threadblock_id_in_cluster,
|
|
warps_per_threadblock_per_core);
|
|
|
|
const float *tile_S = (float *)smem_S;
|
|
#else
|
|
float *tile_S = (float *)arg->addr_q;
|
|
#endif
|
|
|
|
thread_block_online_softmax(tile_S, smem_O, smem_P, tid_in_threadblock,
|
|
threads_per_threadblock, threadblocks_per_cluster,
|
|
threadblock_id_in_cluster, smem_scratchpad,
|
|
smem_rowmax, smem_rowsum);
|
|
|
|
// FIXME unnecessary?
|
|
threadblock_barrier(threadblock_id_in_cluster,
|
|
warps_per_threadblock_per_core);
|
|
|
|
// clear out accumulators
|
|
initialize_accum_regs<0>();
|
|
initialize_accum_regs<1>();
|
|
|
|
load_tile_to_smem<float, MemLayout::MN_major, MemLayout::MN_major, BN, BK>(
|
|
B_COL, 0 /*FIXME*/, 0 /*FIXME*/, gmem_V, smem_V, tid_in_threadblock);
|
|
|
|
threadblock_barrier(threadblock_id_in_cluster,
|
|
warps_per_threadblock_per_core);
|
|
|
|
// FIXME: support MN_major for A for ideal performance
|
|
thread_block_gemm_single_tile<float, MemLayout::K_major, MemLayout::MN_major,
|
|
/*write_to_smem=*/true>(
|
|
smem_P, smem_V, gmem_O /*smem_O*/, tid_in_threadblock,
|
|
threads_per_threadblock, threadblocks_per_cluster,
|
|
threadblock_id_in_cluster);
|
|
|
|
threadblock_barrier(threadblock_id_in_cluster,
|
|
warps_per_threadblock_per_core);
|
|
}
|
|
|
|
int main() {
|
|
kernel_arg_t *arg = (kernel_arg_t *)KERNEL_ARG_DEV_MEM_ADDR;
|
|
|
|
// FIXME:: use actuall seqlen/headdim
|
|
const uint32_t problem_size = (B_ROW * B_COL) / (ELEM_PER_THREAD);
|
|
const uint32_t hw_threads_per_cluster =
|
|
CORES_PER_CLUSTER * vx_num_threads() * vx_num_warps();
|
|
// prevent launching more threads than the necessary problem size
|
|
// TODO: this does not take into account multiple clusters
|
|
const uint32_t grid_size = (problem_size > hw_threads_per_cluster)
|
|
? hw_threads_per_cluster
|
|
: problem_size;
|
|
|
|
#ifdef RADIANCE
|
|
vx_spawn_tasks_cluster(grid_size, (vx_spawn_tasks_cb)kernel_body, arg);
|
|
#else
|
|
// NOTE: This kernel assumes contiguous thread scheduling for efficient shared
|
|
// memory allocation, and therefore does not work with original vx_spawn_tasks
|
|
vx_spawn_tasks_contiguous(grid_size, (vx_spawn_tasks_cb)kernel_body, arg);
|
|
#endif
|
|
return 0;
|
|
}
|