951 lines
36 KiB
C++
951 lines
36 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
|
|
|
|
constexpr uint32_t ROWMAX_SETS = 3;
|
|
constexpr bool DEBUG = true;
|
|
constexpr bool WARP_SPECIALIZED = true;
|
|
|
|
constexpr uint32_t DEV_FAKE_SMEM_START_ADDR = 0xf0000000;
|
|
|
|
// temporary safety stop for wrong configs
|
|
static_assert(NUM_CORES == 4);
|
|
static_assert(NUM_THREADS == 8);
|
|
static_assert(NUM_WARPS == 8);
|
|
|
|
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,
|
|
float *smem_O_row_scale) {
|
|
asm volatile("threadblock_init_sharedmem_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;
|
|
|
|
static_assert((B_ROW % NUM_THREADS) == 0,
|
|
"B_ROW must be a multiple of NUM_THREADS");
|
|
static_assert(B_ROW < (NUM_THREADS * CORES_PER_CLUSTER *
|
|
(NUM_WARPS / (WARP_SPECIALIZED ? 2 : 1))),
|
|
"not enough warps to initialize rowmax/rowsum");
|
|
|
|
// each thread initializes one element in rowmax/rowsum
|
|
// multiple warps participate for the whole vector
|
|
constexpr uint32_t needed_warps = B_ROW / NUM_THREADS;
|
|
if (warp_id < needed_warps /* more warps in HW than needed? */) {
|
|
uint32_t offset = NUM_THREADS * warp_id + tid_in_warp;
|
|
#pragma GCC unroll
|
|
for (int i = 0; i < ROWMAX_SETS; i++) {
|
|
smem_rowmax[offset + i * ROWMAX_SETS] = FLT_MIN;
|
|
}
|
|
smem_rowsum[offset] = 0.0f;
|
|
smem_O_row_scale[offset] = 0.0f;
|
|
}
|
|
|
|
// each warp clears out a row of smem_O
|
|
// FIXME: dedup this pattern
|
|
#pragma GCC unroll 1
|
|
for (int row_offset = 0; row_offset < B_COL;
|
|
row_offset += warps_in_threadblock) {
|
|
const uint32_t row = row_offset + warp_id;
|
|
uint32_t thread_offset = HEADDIM * row + tid_in_warp;
|
|
constexpr uint32_t per_row_iter = HEADDIM / NUM_THREADS;
|
|
const float one = 0.0f;
|
|
#pragma GCC unroll
|
|
for (int i = 0; i < per_row_iter; i++) {
|
|
smem_O[thread_offset] = 0.0f;
|
|
thread_offset += NUM_THREADS;
|
|
}
|
|
}
|
|
|
|
asm volatile("threadblock_init_sharedmem_finish_%=:" ::);
|
|
}
|
|
|
|
inline void thread_block_copy_rowmax(const float *src, float *dest,
|
|
const uint32_t tid_in_threadblock,
|
|
const uint32_t threads_per_threadblock,
|
|
const uint32_t threadblock_id_in_cluster) {
|
|
asm volatile("threadblock_copy_rowmax_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 =
|
|
warps_in_threadblock / CORES_PER_CLUSTER;
|
|
|
|
// each thread copies one element in rowmax
|
|
// multiple warps participate for the whole vector
|
|
constexpr uint32_t num_warps = B_ROW / NUM_THREADS;
|
|
if (warp_id < num_warps) {
|
|
uint32_t offset = NUM_THREADS * warp_id + tid_in_warp;
|
|
dest[offset] = src[offset];
|
|
}
|
|
|
|
threadblock_barrier(threadblock_id_in_cluster,
|
|
warps_per_threadblock_per_core);
|
|
|
|
asm volatile("threadblock_copy_rowmax_finish_%=:" ::);
|
|
}
|
|
|
|
inline void thread_block_copy_tile(const float *src, float *dest,
|
|
const uint32_t tid_in_threadblock,
|
|
const uint32_t threads_per_threadblock,
|
|
const uint32_t threadblock_id_in_cluster) {
|
|
asm volatile("threadblock_copy_tile_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 =
|
|
warps_in_threadblock / CORES_PER_CLUSTER;
|
|
|
|
// FIXME: dedup this pattern
|
|
#pragma GCC unroll 1
|
|
for (int row_offset = 0; row_offset < B_ROW;
|
|
row_offset += warps_in_threadblock) {
|
|
const uint32_t row = row_offset + warp_id;
|
|
const uint32_t first_thread_offset = B_COL * row;
|
|
|
|
constexpr uint32_t per_row_iter = B_COL / NUM_THREADS;
|
|
uint32_t thread_offset = first_thread_offset + tid_in_warp;
|
|
#pragma GCC unroll
|
|
for (int i = 0; i < per_row_iter; i++) {
|
|
dest[thread_offset] = src[thread_offset];
|
|
thread_offset += NUM_THREADS;
|
|
}
|
|
|
|
threadblock_barrier(threadblock_id_in_cluster,
|
|
warps_per_threadblock_per_core);
|
|
}
|
|
|
|
asm volatile("threadblock_copy_tile_finish_%=:" ::);
|
|
}
|
|
|
|
template <int order>
|
|
inline float exponential_taylor_term(const float x) {
|
|
asm volatile("exponential_taylor_term_start_%=:" ::);
|
|
|
|
float res = 1.0f;
|
|
|
|
if constexpr (order == 1) {
|
|
res = x;
|
|
} else if constexpr (order == 2) {
|
|
res = x * x;
|
|
res /= 2.0f;
|
|
} else if constexpr (order == 3) {
|
|
res = x * x * x;
|
|
res /= 6.0f;
|
|
}
|
|
|
|
asm volatile("exponential_taylor_term_end_%=:" ::);
|
|
return res;
|
|
}
|
|
|
|
__attribute__((always_inline)) inline void thread_block_online_softmax(
|
|
const float *smem_S, float *smem_P, const uint32_t tid_in_threadblock,
|
|
const uint32_t threads_per_threadblock,
|
|
const uint32_t threadblock_id_in_cluster, float *smem_scratchpad,
|
|
float *smem_rowmax, float *smem_rowsum, float *smem_O_row_scale) {
|
|
asm volatile("thread_block_online_softmax_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 =
|
|
warps_in_threadblock / CORES_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]));
|
|
|
|
float *smem_rowmax_this = smem_rowmax + B_ROW;
|
|
|
|
#pragma GCC unroll 1
|
|
for (int row_offset = 0; row_offset < B_ROW;
|
|
row_offset += warps_in_threadblock) {
|
|
const uint32_t row = row_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 down to one row max
|
|
// one warp handles one row in tile
|
|
|
|
constexpr uint32_t per_row_iter = B_COL / NUM_THREADS;
|
|
uint32_t thread_offset = first_thread_offset + tid_in_warp;
|
|
// FIXME: threadblock_id needs to be in here too
|
|
float *warp_smem = smem_scratchpad + (warp_id * NUM_THREADS);
|
|
|
|
// #define DUMB_ROWMAX
|
|
#ifdef DUMB_ROWMAX
|
|
// FIXME remove
|
|
threadblock_barrier(threadblock_id_in_cluster,
|
|
warps_per_threadblock_per_core);
|
|
|
|
// no tree reduction; a single thread in a warp does serialized max across
|
|
// the entire row
|
|
if (tid_in_warp == 0) {
|
|
float rowmax = smem_S[first_thread_offset];
|
|
#pragma GCC unroll 16
|
|
for (int i = 0; i < B_COL; i++) {
|
|
asm volatile("fmax.s %0, %1, %2"
|
|
: "=f"(rowmax)
|
|
: "f"(rowmax), "f"(smem_S[first_thread_offset + i]));
|
|
}
|
|
smem_rowmax_this[row] = rowmax;
|
|
|
|
// update previous rowmax
|
|
// i.e. mi_new = max(mi, mij)
|
|
float prev_rowmax = smem_rowmax[row];
|
|
// stage prev rowmax in scratchpad for warp-wide broadcast
|
|
warp_smem[0] = prev_rowmax;
|
|
asm volatile("fmax.s %0, %1, %2"
|
|
: "=f"(rowmax)
|
|
: "f"(rowmax), "f"(prev_rowmax));
|
|
smem_rowmax[row] = rowmax;
|
|
}
|
|
|
|
#else
|
|
static_assert((B_COL % NUM_THREADS) == 0,
|
|
"B_COL must be a multiple of NUM_THREADS");
|
|
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
|
|
warp_smem[tid_in_warp] = per_thread_max;
|
|
|
|
// sync writes to warp_smem
|
|
threadblock_barrier(threadblock_id_in_cluster,
|
|
warps_per_threadblock_per_core);
|
|
|
|
// #define PARALLEL_ROWMAX
|
|
#ifndef PARALLEL_ROWMAX
|
|
// 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 i = 1; i < NUM_THREADS; i++) {
|
|
float other = warp_smem[i];
|
|
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[row];
|
|
// stage prev rowmax in scratchpad for warp-wide broadcast
|
|
warp_smem[0] = prev_rowmax;
|
|
asm volatile("fmax.s %0, %1, %2"
|
|
: "=f"(rowmax)
|
|
: "f"(rowmax), "f"(prev_rowmax));
|
|
smem_rowmax[row] = rowmax;
|
|
}
|
|
#else
|
|
if (warp_id < warps_in_threadblock / NUM_THREADS) {
|
|
const uint32_t row = row_offset + NUM_THREADS * warp_id + tid_in_warp;
|
|
float *const thread_smem = smem_scratchpad + (tid_in_warp * NUM_THREADS);
|
|
float rowmax = FLT_MIN;
|
|
#pragma GCC unroll
|
|
for (int i = 0; i < NUM_THREADS; i++) {
|
|
const float f = thread_smem[i];
|
|
asm volatile("fmax.s %0, %1, %2" : "=f"(rowmax) : "f"(rowmax), "f"(f));
|
|
}
|
|
smem_rowmax_this[row] = rowmax;
|
|
|
|
// update previous rowmax
|
|
// i.e. mi_new = max(mi, mij)
|
|
float prev_rowmax = smem_rowmax[row];
|
|
// stage prev rowmax in scratchpad for warp-wide broadcast
|
|
thread_smem[0] = prev_rowmax;
|
|
asm volatile("fmax.s %0, %1, %2"
|
|
: "=f"(rowmax)
|
|
: "f"(rowmax), "f"(prev_rowmax));
|
|
smem_rowmax[row] = rowmax;
|
|
}
|
|
#endif // PARALLEL_ROWMAX
|
|
#endif // DUMB_ROWMAX
|
|
|
|
threadblock_barrier(threadblock_id_in_cluster,
|
|
warps_per_threadblock_per_core);
|
|
|
|
// broadcast prev rowmax to all threads in the warp
|
|
// NOTE: memory consistency is a little sketchy here
|
|
const float rowmax_prev = warp_smem[0];
|
|
const float rowmax_this = smem_rowmax_this[row];
|
|
|
|
// 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 updated rowmax to all threads in the warp
|
|
const float rowmax_new = smem_rowmax[row];
|
|
|
|
asm volatile("flashattn_exp_p_start_%=:" ::);
|
|
|
|
thread_offset = first_thread_offset + tid_in_warp;
|
|
#pragma GCC unroll
|
|
for (int i = 0; i < per_row_iter; i++) {
|
|
float f0 = smem_S[thread_offset];
|
|
|
|
f0 -= rowmax_new;
|
|
|
|
// 2nd-order Taylor approximation
|
|
float exp = 1.0f;
|
|
exp += exponential_taylor_term<1>(f0);
|
|
exp += exponential_taylor_term<2>(f0);
|
|
|
|
// Store S transposed to the shared memory
|
|
|
|
smem_P[thread_offset] = exp;
|
|
|
|
thread_offset += NUM_THREADS;
|
|
}
|
|
|
|
asm volatile("flashattn_exp_p_end_%=:" ::);
|
|
|
|
|
|
threadblock_barrier(threadblock_id_in_cluster,
|
|
warps_per_threadblock_per_core);
|
|
|
|
// rowsum
|
|
//
|
|
// two-level tree reduction, similar to rowmax
|
|
|
|
asm volatile("flashattn_rowsum_start_%=:" ::);
|
|
|
|
float per_thread_sum = 0.0f;
|
|
|
|
thread_offset = first_thread_offset + tid_in_warp;
|
|
#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 = rowmax_prev;
|
|
const float mi_this = rowmax_this;
|
|
|
|
const float x = mi_prev - mi_this;
|
|
// 2nd-order Taylor approximation
|
|
float exp = 1.0f;
|
|
exp += exponential_taylor_term<1>(x);
|
|
exp += exponential_taylor_term<2>(x);
|
|
|
|
// update rowsum
|
|
const float rowsum_prev = smem_rowsum[row];
|
|
float rowsum_new = exp * rowsum_prev + rowsum;
|
|
|
|
smem_rowsum[row] = rowsum_new;
|
|
}
|
|
|
|
asm volatile("flashattn_rowsum_end_%=:" ::);
|
|
|
|
threadblock_barrier(threadblock_id_in_cluster,
|
|
warps_per_threadblock_per_core);
|
|
|
|
// compute Oi rescale factor
|
|
// FIXME: parallelize this across threads
|
|
//
|
|
asm volatile("flashattn_rescale_factor_start_%=:" ::);
|
|
|
|
thread_offset = first_thread_offset + tid_in_warp;
|
|
#pragma GCC unroll
|
|
for (int i = 0; i < per_row_iter; i++) {
|
|
const float mi_prev = rowmax_prev;
|
|
const float mi_new = rowmax_new;
|
|
|
|
const float x = mi_prev - mi_new;
|
|
// 2nd-order Taylor approximation
|
|
float exp = 1.0f;
|
|
exp += exponential_taylor_term<1>(x);
|
|
exp += exponential_taylor_term<2>(x);
|
|
|
|
// @perf: div vs. expansion on e(-x)?
|
|
smem_O_row_scale[row] = 1.0f / exp;
|
|
|
|
thread_offset += NUM_THREADS;
|
|
}
|
|
|
|
asm volatile("flashattn_rescale_factor_end_%=:" ::);
|
|
|
|
threadblock_barrier(threadblock_id_in_cluster,
|
|
warps_per_threadblock_per_core);
|
|
}
|
|
|
|
asm volatile("thread_block_online_softmax_finish_%=:" ::);
|
|
}
|
|
|
|
__attribute__((always_inline)) inline void thread_block_O_rescale(
|
|
const float *smem_O_in, float *smem_O_out, const float *smem_O_row_scale,
|
|
const uint32_t tid_in_threadblock, const uint32_t threads_per_threadblock,
|
|
const uint32_t threadblock_id_in_cluster) {
|
|
asm volatile("thread_block_O_rescale_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 =
|
|
warps_in_threadblock / CORES_PER_CLUSTER;
|
|
|
|
#pragma GCC unroll 1
|
|
for (int row_offset = 0; row_offset < B_ROW;
|
|
row_offset += warps_in_threadblock) {
|
|
const uint32_t row = row_offset + warp_id;
|
|
const uint32_t first_thread_offset = B_COL * row;
|
|
constexpr uint32_t per_row_iter = B_COL / NUM_THREADS;
|
|
uint32_t thread_offset = first_thread_offset + tid_in_warp;
|
|
|
|
// Oi rescale
|
|
//
|
|
#pragma GCC unroll
|
|
for (int i = 0; i < per_row_iter; i++) {
|
|
const float o = smem_O_in[thread_offset];
|
|
const float scale = smem_O_row_scale[row];
|
|
smem_O_out[thread_offset] = (o * scale);
|
|
|
|
thread_offset += NUM_THREADS;
|
|
}
|
|
}
|
|
|
|
asm volatile("thread_block_O_rescale_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
|
|
constexpr uint32_t threads_per_threadblock_theoretical =
|
|
(B_ROW * B_COL) / (ELEM_PER_THREAD);
|
|
constexpr uint32_t hw_threads_per_cluster =
|
|
CORES_PER_CLUSTER * NUM_THREADS * NUM_WARPS;
|
|
// cap maximum threadblock size to # of HW threads in cluster, to prevent
|
|
// multiple "wave" invocations which slows down the kernel
|
|
constexpr uint32_t threads_per_threadblock =
|
|
(threads_per_threadblock_theoretical > hw_threads_per_cluster)
|
|
? hw_threads_per_cluster
|
|
: threads_per_threadblock_theoretical;
|
|
constexpr uint32_t threadblocks_per_cluster =
|
|
hw_threads_per_cluster / threads_per_threadblock;
|
|
constexpr uint32_t warps_per_threadblock_per_core =
|
|
NUM_WARPS / threadblocks_per_cluster;
|
|
|
|
const uint32_t threadblock_id = task_id / threads_per_threadblock;
|
|
const uint32_t threadblock_id_in_cluster =
|
|
threadblock_id % threadblocks_per_cluster;
|
|
const uint32_t tid_in_threadblock = task_id % threads_per_threadblock;
|
|
const uint32_t warp_id = tid_in_threadblock / NUM_THREADS;
|
|
constexpr uint32_t warps_in_threadblock =
|
|
threads_per_threadblock / NUM_THREADS;
|
|
|
|
// warpgroup context
|
|
constexpr uint32_t threads_per_warpgroup =
|
|
threads_per_threadblock / (WARP_SPECIALIZED ? 2 : 1);
|
|
constexpr uint32_t warpgroups_per_cluster =
|
|
threadblocks_per_cluster * (WARP_SPECIALIZED ? 2 : 1);
|
|
const uint32_t warps_per_warpgroup_per_core =
|
|
NUM_WARPS / warpgroups_per_cluster;
|
|
const uint32_t warpgroup_id = task_id / threads_per_warpgroup;
|
|
const uint32_t warpgroup_id_in_cluster =
|
|
warpgroup_id % warpgroups_per_cluster;
|
|
const uint32_t tid_in_warpgroup = tid_in_threadblock % threads_per_warpgroup;
|
|
|
|
// FIXME do proper software pipelining
|
|
// if (WARP_SPECIALIZED && warpgroup_id_in_cluster != 1) {
|
|
// return;
|
|
// }
|
|
|
|
const uint32_t dim_seqlen = arg->dim_seqlen;
|
|
const uint32_t dim_headdim = arg->dim_headdim;
|
|
|
|
// get global memory addresses from kernel arguments
|
|
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_tmp_d0 = reinterpret_cast<float *>(0xd0000000UL);
|
|
float *gmem_tmp_d1 = reinterpret_cast<float *>(0xd1000000UL);
|
|
float *gmem_tmp_d2 = reinterpret_cast<float *>(0xd2000000UL);
|
|
float *gmem_tmp_d3 = reinterpret_cast<float *>(0xd3000000UL);
|
|
float *gmem_tmp_d4 = reinterpret_cast<float *>(0xd4000000UL);
|
|
float *gmem_tmp_d5 = reinterpret_cast<float *>(0xd5000000UL);
|
|
float *gmem_tmp_d6 = reinterpret_cast<float *>(0xd6000000UL);
|
|
float *gmem_tmp_d7 = reinterpret_cast<float *>(0xd7000000UL);
|
|
float *gmem_tmp_e0 = reinterpret_cast<float *>(0xe0000000UL);
|
|
float *gmem_tmp_e1 = reinterpret_cast<float *>(0xe1000000UL);
|
|
float *gmem_tmp_e2 = reinterpret_cast<float *>(0xe2000000UL);
|
|
float *gmem_tmp_e3 = reinterpret_cast<float *>(0xe3000000UL);
|
|
|
|
// static shared memory allocation
|
|
constexpr uint32_t smem_Q_size = B_ROW * HEADDIM;
|
|
constexpr uint32_t smem_K_size = B_COL * 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_cursor = reinterpret_cast<float *>(DEV_FAKE_SMEM_START_ADDR);
|
|
float *smem_Q0 = smem_cursor;
|
|
smem_cursor += smem_Q_size;
|
|
float *smem_Q1 = smem_cursor;
|
|
smem_cursor += smem_Q_size;
|
|
float *smem_K0 = smem_cursor;
|
|
smem_cursor += smem_K_size;
|
|
float *smem_K1 = smem_cursor;
|
|
smem_cursor += smem_K_size;
|
|
float *smem_V0 = smem_cursor;
|
|
smem_cursor += smem_V_size;
|
|
float *smem_V1 = smem_cursor;
|
|
smem_cursor += smem_V_size;
|
|
float *smem_S0 = smem_cursor;
|
|
smem_cursor += smem_QK_size;
|
|
float *smem_S1 = smem_cursor;
|
|
smem_cursor += smem_QK_size;
|
|
float *smem_P0 = smem_S0; // in-place update
|
|
float *smem_P1 = smem_S1; // in-place update
|
|
float *smem_O0 = smem_cursor;
|
|
smem_cursor += smem_O_size;
|
|
float *smem_O1 = smem_cursor;
|
|
smem_cursor += smem_O_size;
|
|
|
|
// allocate rowmax/rowsum storage at the end of the sharedmem address space
|
|
constexpr uint32_t smem_rowmax_size = B_ROW * ROWMAX_SETS;
|
|
constexpr uint32_t smem_rowsum_size = B_ROW;
|
|
constexpr uint32_t smem_O_row_scale_size = B_ROW;
|
|
smem_cursor = reinterpret_cast<float *>(SMEM_ADDR_END);
|
|
|
|
smem_cursor -= smem_rowmax_size;
|
|
float *smem_rowmax_0 = smem_cursor;
|
|
smem_cursor -= smem_rowmax_size;
|
|
float *smem_rowmax_1 = smem_cursor;
|
|
smem_cursor -= smem_rowsum_size;
|
|
float *smem_rowsum_0 = smem_cursor;
|
|
smem_cursor -= smem_rowsum_size;
|
|
float *smem_rowsum_1 = smem_cursor;
|
|
smem_cursor -= smem_O_row_scale_size;
|
|
float *smem_O_row_scale_0 = smem_cursor;
|
|
smem_cursor -= smem_O_row_scale_size;
|
|
float *smem_O_row_scale_1 = smem_cursor;
|
|
|
|
// sharedmem "scratchpad" area to put temporary data, e.g. for tree reduction
|
|
// in rowsum
|
|
// NOTE: out-of bounds is not checked
|
|
// TODO: reduce this from B_ROW to NUM_WARPS
|
|
constexpr uint32_t smem_scratchpad_size =
|
|
threads_per_warpgroup * 2 /*arbitrary slack*/;
|
|
smem_cursor -= smem_scratchpad_size;
|
|
float *smem_scratchpad_0 = smem_cursor;
|
|
smem_cursor -= smem_scratchpad_size;
|
|
float *smem_scratchpad_1 = smem_cursor;
|
|
|
|
// select the correct buffer by warpgroup
|
|
float *smem_Q = (warpgroup_id % 2) ? smem_Q1 : smem_Q0;
|
|
float *smem_K = (warpgroup_id % 2) ? smem_K1 : smem_K0;
|
|
float *smem_V = (warpgroup_id % 2) ? smem_V1 : smem_V0;
|
|
float *smem_S = (warpgroup_id % 2) ? smem_S1 : smem_S0;
|
|
float *smem_O = (warpgroup_id % 2) ? smem_O1 : smem_O0;
|
|
float *smem_P = smem_S;
|
|
float *smem_O_row_scale =
|
|
(warpgroup_id % 2) ? smem_O_row_scale_1 : smem_O_row_scale_0;
|
|
float *smem_rowmax = (warpgroup_id % 2) ? smem_rowmax_1 : smem_rowmax_0;
|
|
float *smem_rowsum = (warpgroup_id % 2) ? smem_rowsum_1 : smem_rowsum_0;
|
|
float *smem_scratchpad =
|
|
(warpgroup_id % 2) ? smem_scratchpad_1 : smem_scratchpad_0;
|
|
|
|
// initialize rowmax/rowsum values in sharedmem
|
|
thread_block_init_sharedmem(tid_in_warpgroup, threads_per_warpgroup, smem_O,
|
|
smem_rowmax, smem_rowsum, smem_O_row_scale);
|
|
|
|
constexpr uint32_t global_barrier_id = NUM_WARPS - 1; // arbitrary
|
|
|
|
// delay warpgroup 0 by 1 iteration to do ping-pong scheduling
|
|
if (warpgroup_id == 1) {
|
|
threadblock_barrier(global_barrier_id, warps_per_threadblock_per_core);
|
|
}
|
|
|
|
// read Q and K into SMEM before the loop starts
|
|
//
|
|
static_assert(B_ROW == B_COL, "currently only supports square tiles");
|
|
|
|
// load Q; this stays in SMEM for the entire loop
|
|
load_tile_to_smem<float, MemLayout::MN_major, MemLayout::MN_major, B_ROW,
|
|
HEADDIM, threads_per_warpgroup>(
|
|
dim_seqlen, warpgroup_id, 0 /* dim_k == headdim */, gmem_Q, smem_Q,
|
|
tid_in_warpgroup);
|
|
|
|
// load K
|
|
load_tile_to_smem<float, MemLayout::MN_major, MemLayout::MN_major, B_COL,
|
|
HEADDIM, threads_per_warpgroup>(
|
|
dim_seqlen, /*tile_k=*/0, 0 /* dim_k == headdim */, gmem_K, smem_K,
|
|
tid_in_warpgroup);
|
|
|
|
// protect write to SMEM
|
|
threadblock_barrier(warpgroup_id_in_cluster, warps_per_warpgroup_per_core);
|
|
|
|
asm volatile ("tile_loop_start_%=:" :: );
|
|
|
|
// "inner loop" along the columns of K^T
|
|
const uint32_t k_tiles = (dim_seqlen / B_COL);
|
|
for (uint32_t tile_k = 0; tile_k < k_tiles; tile_k++) {
|
|
// float *smem_P_produce = (tile_k % 2) ? smem_P0 : smem_P1;
|
|
// float *smem_P_consume = (tile_k % 2) ? smem_P1 : smem_P0;
|
|
// float *smem_V_produce = (tile_k % 2) ? smem_V0 : smem_V1;
|
|
// float *smem_V_consume = (tile_k % 2) ? smem_V1 : smem_V0;
|
|
// float *smem_O_row_scale_produce =
|
|
// (tile_k % 2) ? smem_O_row_scale_0 : smem_O_row_scale_1;
|
|
// float *smem_O_row_scale_consume =
|
|
// (tile_k % 2) ? smem_O_row_scale_1 : smem_O_row_scale_0;
|
|
|
|
constexpr bool skip_gemm_qk = false;
|
|
if constexpr (!skip_gemm_qk) {
|
|
// GEMM I: S = Q*K
|
|
//
|
|
// FIXME: deduplicate this between GEMM II
|
|
if constexpr (!WARP_SPECIALIZED) {
|
|
// clear out accumulators before GEMM
|
|
initialize_accum_regs<0>();
|
|
initialize_accum_regs<1>();
|
|
|
|
thread_block_gemm_single_tile<
|
|
float, MemLayout::MN_major, MemLayout::MN_major, B_ROW, B_COL,
|
|
HEADDIM, /*leading_dim_a=*/0, /*leading_dim_b=*/0,
|
|
/*load_accum=*/false,
|
|
/*write_to_smem=*/true>(
|
|
smem_Q, smem_K, nullptr /*ignore accum*/, smem_S, tid_in_warpgroup,
|
|
threads_per_warpgroup, warpgroups_per_cluster,
|
|
warpgroup_id_in_cluster);
|
|
} else {
|
|
// when warp-specialized, there's only enough warps to do 64x32 tile
|
|
// size so we need to do 2 GEMM calls
|
|
static_assert(B_ROW / 2 == 32,
|
|
"tile size assumption for warp-specialization not met");
|
|
|
|
// assumes smem_Q is K-major
|
|
// FIXME: fix this to MN-major
|
|
float *smem_Q_half0 = smem_Q;
|
|
float *smem_Q_half1 = smem_Q + (B_ROW / 2); // MN-major
|
|
// float *smem_Q_half1 = smem_Q + (B_ROW / 2) * HEADDIM; // K-major
|
|
float *smem_S_half0 = smem_S;
|
|
float *smem_S_half1 = smem_S + (B_ROW / 2) * B_COL;
|
|
|
|
// clear out accumulators before GEMM
|
|
initialize_accum_regs<0>();
|
|
initialize_accum_regs<1>();
|
|
|
|
// split by rows into 2 chunks
|
|
thread_block_gemm_single_tile<
|
|
float, MemLayout::MN_major, MemLayout::MN_major, B_ROW / 2,
|
|
B_COL, HEADDIM, /*leading_dim_a=*/B_ROW, /*leading_dim_b=*/0,
|
|
/*load_accum=*/false,
|
|
/*write_to_smem=*/true>(
|
|
smem_Q_half0, smem_K, nullptr /*ignore accum*/, smem_S_half0,
|
|
tid_in_warpgroup, threads_per_warpgroup, warpgroups_per_cluster,
|
|
warpgroup_id_in_cluster);
|
|
|
|
initialize_accum_regs<0>();
|
|
initialize_accum_regs<1>();
|
|
|
|
thread_block_gemm_single_tile<
|
|
float, MemLayout::MN_major, MemLayout::MN_major, B_ROW / 2,
|
|
B_COL, HEADDIM, /*leading_dim_a=*/B_ROW, /*leading_dim_b=*/0,
|
|
/*load_accum=*/false,
|
|
/*write_to_smem=*/true>(
|
|
smem_Q_half1, smem_K, nullptr /*ignore accum*/, smem_S_half1,
|
|
tid_in_warpgroup, threads_per_warpgroup, warpgroups_per_cluster,
|
|
warpgroup_id_in_cluster);
|
|
}
|
|
} else {
|
|
// load Q*K
|
|
load_tile_to_smem<float, MemLayout::K_major, MemLayout::K_major, B_COL,
|
|
HEADDIM, threads_per_warpgroup>(
|
|
dim_seqlen, warpgroup_id /* parallelize across rows */, tile_k,
|
|
gmem_Q /*contains S*/, smem_S, tid_in_warpgroup);
|
|
}
|
|
|
|
// protect write to SMEM (smem_S) before softmax
|
|
threadblock_barrier(warpgroup_id_in_cluster, warps_per_warpgroup_per_core);
|
|
|
|
if constexpr (DEBUG) {
|
|
if (warpgroup_id == 0) {
|
|
if (tile_k == 0) {
|
|
thread_block_copy_tile(smem_S, gmem_tmp_d0,
|
|
tid_in_warpgroup, threads_per_warpgroup,
|
|
warpgroup_id_in_cluster);
|
|
} else if (tile_k == 1) {
|
|
thread_block_copy_tile(smem_S, gmem_tmp_d1,
|
|
tid_in_warpgroup, threads_per_warpgroup,
|
|
warpgroup_id_in_cluster);
|
|
}
|
|
|
|
threadblock_barrier(warpgroup_id_in_cluster,
|
|
warps_per_warpgroup_per_core);
|
|
}
|
|
}
|
|
|
|
// inter-warpgroup barrier before online softmax
|
|
threadblock_barrier(global_barrier_id, warps_per_threadblock_per_core);
|
|
|
|
// Online softmax
|
|
//
|
|
thread_block_online_softmax(smem_S, smem_P, tid_in_warpgroup,
|
|
threads_per_warpgroup, warpgroup_id_in_cluster,
|
|
smem_scratchpad, smem_rowmax, smem_rowsum,
|
|
smem_O_row_scale);
|
|
|
|
// data movement for K and V
|
|
//
|
|
// Q stays in SMEM for the entire loop
|
|
//
|
|
// load K for the next iteration
|
|
load_tile_to_smem<float, MemLayout::MN_major, MemLayout::MN_major, B_COL,
|
|
HEADDIM, threads_per_warpgroup>(
|
|
dim_seqlen, tile_k + 1, 0 /* dim_k == headdim */, gmem_K, smem_K,
|
|
tid_in_warpgroup);
|
|
|
|
// load V for the current iteration
|
|
// V dimension is [seqlen, headdim], stored N(headdim)-major
|
|
load_tile_to_smem<float, MemLayout::MN_major, MemLayout::MN_major, B_COL,
|
|
HEADDIM, threads_per_warpgroup>(
|
|
HEADDIM, 0 /* full N-dimension */, tile_k, gmem_V, smem_V,
|
|
tid_in_warpgroup);
|
|
|
|
// protect write to SMEM
|
|
threadblock_barrier(warpgroup_id_in_cluster, warps_per_warpgroup_per_core);
|
|
|
|
if constexpr (DEBUG) {
|
|
if (warpgroup_id == 0) {
|
|
if (tile_k == 0) {
|
|
thread_block_copy_rowmax(smem_rowmax, gmem_tmp_e0, tid_in_warpgroup,
|
|
threads_per_warpgroup,
|
|
warpgroup_id_in_cluster);
|
|
thread_block_copy_rowmax(smem_rowsum, gmem_tmp_e2, tid_in_warpgroup,
|
|
threads_per_warpgroup,
|
|
warpgroup_id_in_cluster);
|
|
} else if (tile_k == 1) {
|
|
thread_block_copy_rowmax(smem_rowmax, gmem_tmp_e1, tid_in_warpgroup,
|
|
threads_per_warpgroup,
|
|
warpgroup_id_in_cluster);
|
|
thread_block_copy_rowmax(smem_rowsum, gmem_tmp_e3, tid_in_warpgroup,
|
|
threads_per_warpgroup,
|
|
warpgroup_id_in_cluster);
|
|
}
|
|
|
|
threadblock_barrier(warpgroup_id_in_cluster,
|
|
warps_per_warpgroup_per_core);
|
|
}
|
|
}
|
|
|
|
// inter-warpgroup barrier before GEMM II
|
|
threadblock_barrier(global_barrier_id, warps_per_threadblock_per_core);
|
|
|
|
// GEMM II: O = O + P*V
|
|
|
|
// Oi rescale
|
|
thread_block_O_rescale(smem_O, smem_O /*in-place*/,
|
|
smem_O_row_scale, tid_in_warpgroup,
|
|
threads_per_warpgroup, warpgroup_id_in_cluster);
|
|
|
|
// rescale-to-PV-GEMM barrier
|
|
threadblock_barrier(warpgroup_id_in_cluster, warps_per_warpgroup_per_core);
|
|
|
|
if constexpr (DEBUG) {
|
|
if (warpgroup_id == 0) {
|
|
// O before PV
|
|
if (tile_k == 0) {
|
|
thread_block_copy_tile(smem_P, gmem_tmp_d2, tid_in_warpgroup,
|
|
threads_per_warpgroup,
|
|
warpgroup_id_in_cluster);
|
|
thread_block_copy_tile(smem_O, gmem_tmp_d4, tid_in_warpgroup,
|
|
threads_per_warpgroup,
|
|
warpgroup_id_in_cluster);
|
|
} else if (tile_k == 1) {
|
|
thread_block_copy_tile(smem_P, gmem_tmp_d3, tid_in_warpgroup,
|
|
threads_per_warpgroup,
|
|
warpgroup_id_in_cluster);
|
|
thread_block_copy_tile(smem_O, gmem_tmp_d5, tid_in_warpgroup,
|
|
threads_per_warpgroup,
|
|
warpgroup_id_in_cluster);
|
|
}
|
|
|
|
threadblock_barrier(warpgroup_id_in_cluster,
|
|
warps_per_warpgroup_per_core);
|
|
}
|
|
}
|
|
|
|
if constexpr (!WARP_SPECIALIZED) {
|
|
// clear out accumulators before GEMM
|
|
initialize_accum_regs<0>();
|
|
initialize_accum_regs<1>();
|
|
|
|
thread_block_gemm_single_tile<float, MemLayout::K_major,
|
|
MemLayout::MN_major, B_ROW, HEADDIM, B_COL,
|
|
/*leading_dim_a=*/0, /*leading_dim_b=*/0,
|
|
/*load_accum=*/true,
|
|
/*write_to_smem=*/true>(
|
|
smem_P, smem_V, smem_O /*load accum*/, smem_O, tid_in_warpgroup,
|
|
threads_per_warpgroup, warpgroups_per_cluster,
|
|
warpgroup_id_in_cluster);
|
|
|
|
// FIXME: wrong but fast
|
|
// thread_block_gemm_single_tile<float, MemLayout::MN_major,
|
|
// MemLayout::MN_major,
|
|
// B_ROW, HEADDIM, B_COL,
|
|
// /*leading_dim_a=*/0, /*leading_dim_b=*/0,
|
|
// /*load_accum=*/true,
|
|
// /*write_to_smem=*/true>(
|
|
// smem_P, smem_V, smem_O /*load accum*/, smem_O,
|
|
// tid_in_warpgroup, threads_per_warpgroup,
|
|
// warpgroups_per_cluster, warpgroup_id_in_cluster);
|
|
} else {
|
|
// when warp-specialized, there's only enough warps to do 64x32 tile
|
|
// size so we need to do 2 GEMM calls
|
|
static_assert(B_ROW / 2 == 32,
|
|
"tile size assumption for warp-specialization not met");
|
|
|
|
// assumes smem_P is K-major
|
|
float *smem_P_half0 = smem_P;
|
|
float *smem_P_half1 = smem_P + (B_ROW / 2) * B_COL;
|
|
float *smem_O_half0 = smem_O;
|
|
float *smem_O_half1 = smem_O + (B_ROW / 2) * HEADDIM;
|
|
|
|
// clear out accumulators before GEMM
|
|
initialize_accum_regs<0>();
|
|
initialize_accum_regs<1>();
|
|
|
|
// split by rows into 2 chunks
|
|
thread_block_gemm_single_tile<
|
|
float, MemLayout::K_major, MemLayout::MN_major, B_ROW / 2, HEADDIM,
|
|
B_COL, /*leading_dim_a=*/0, /*leading_dim_b=*/0,
|
|
/*load_accum=*/true,
|
|
/*write_to_smem=*/true>(
|
|
smem_P_half0, smem_V, smem_O_half0 /*load accum*/, smem_O_half0,
|
|
tid_in_warpgroup, threads_per_warpgroup, warpgroups_per_cluster,
|
|
warpgroup_id_in_cluster);
|
|
|
|
initialize_accum_regs<0>();
|
|
initialize_accum_regs<1>();
|
|
|
|
thread_block_gemm_single_tile<
|
|
float, MemLayout::K_major, MemLayout::MN_major, B_ROW / 2, HEADDIM,
|
|
B_COL, /*leading_dim_a=*/0, /*leading_dim_b=*/0,
|
|
/*load_accum=*/true,
|
|
/*write_to_smem=*/true>(
|
|
smem_P_half1, smem_V, smem_O_half1 /*load accum*/, smem_O_half1,
|
|
tid_in_warpgroup, threads_per_warpgroup, warpgroups_per_cluster,
|
|
warpgroup_id_in_cluster);
|
|
}
|
|
|
|
threadblock_barrier(warpgroup_id_in_cluster, warps_per_warpgroup_per_core);
|
|
|
|
if constexpr (DEBUG) {
|
|
if (warpgroup_id == 0) {
|
|
// O after PV
|
|
if (tile_k == 0) {
|
|
thread_block_copy_tile(smem_O, gmem_tmp_d6, tid_in_warpgroup,
|
|
threads_per_warpgroup,
|
|
warpgroup_id_in_cluster);
|
|
} else if (tile_k == 1) {
|
|
thread_block_copy_tile(smem_O, gmem_tmp_d7, tid_in_warpgroup,
|
|
threads_per_warpgroup,
|
|
warpgroup_id_in_cluster);
|
|
}
|
|
|
|
threadblock_barrier(warpgroup_id_in_cluster,
|
|
warps_per_warpgroup_per_core);
|
|
}
|
|
}
|
|
|
|
tile_iter_end:
|
|
// synchronize progress of two warpgroups
|
|
// threadblock_barrier(threadblock_id_in_cluster,
|
|
// warps_per_threadblock_per_core);
|
|
// threadblock_barrier(3, // FIXME
|
|
// NUM_WARPS);
|
|
}
|
|
|
|
asm volatile ("tile_loop_finish_%=:" :: );
|
|
|
|
// wait for warpgroup 1 to finish, which called the global barrier before
|
|
// entering the loop
|
|
if (warpgroup_id == 0) {
|
|
threadblock_barrier(global_barrier_id, 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;
|
|
}
|