sgemm_tcore: Add GEMMINI_DMA to non-warp-specialized mode

~63% util for 128x128; ~83% for the k-loop.

FIXME: result is not correct currently. Need to fix the transpose
This commit is contained in:
Hansung Kim
2024-06-10 17:06:46 -07:00
parent 51d9cffb2d
commit a22762db94

View File

@@ -5,9 +5,23 @@
#include <vx_print.h>
#include <vx_spawn.h>
#include "common.h"
#include "include/gemmini.h"
#include "gemmini_mmio.h"
#define NUM_LANES 8
#if SMEM_SIZE != 0x4000
#error Currently only supports 16K spad
#endif
#define SMEM_ADDR_Q0 ((float * const) 0xff000000)
#define SMEM_ADDR_Q1 ((float * const) 0xff001000)
#define SMEM_ADDR_Q2 ((float * const) 0xff002000)
#define SMEM_ADDR_Q3 ((float * const) 0xff003000)
#define SPAD_ADDR_Q0 0x0
#define SPAD_ADDR_Q1 0x80
#define SPAD_ADDR_Q2 0x100
#define SPAD_ADDR_Q3 0x180
// number of loop around the inner 0..TCK..BK loop to simulate perfect-DRAM
// scenario
#define BK_LOOP 1
@@ -16,8 +30,7 @@
// 1: GMEM loads of A matrix
// 0: SMEM stores of A matrix
#define GMEM_COALESCED_A 1
#define DOUBLE_BUFFER 1
#define GEMMINI_DMA 1
// Constraints on parameters:
// * Memory:
@@ -41,7 +54,7 @@
#define TCK 8
#define WMITER (WM / TCM)
#define WNITER (WN / TCN)
#define ELEM_PER_THREAD (WMITER * WNITER * ((TCM * TCN) / NUM_LANES) / (DOUBLE_BUFFER ? 2 : 1))
#define ELEM_PER_THREAD (WMITER * WNITER * (TCM * TCN) / NUM_LANES)
// FIXME: NUM_THREADS and NUM_WARPS hardcoded
#if ((BM * BN / ELEM_PER_THREAD) > (CORES_PER_CLUSTER * 8 * 8))
@@ -262,15 +275,16 @@ inline void write_results(const int thread_in_warp, const int warp_col,
float *C, const int threadblock_id_x,
const int threadblock_id_y) {
int tid = thread_in_warp;
int tg = tid / 4;
// these are [0, TCM/TCN)
int tid_row = 0;
int tid_col = 0;
map_c(tid, tid_row, tid_col);
int local_row = (WM * warp_row + TCM * wm_iter) + tid_row;
int local_col = (WN * warp_col + TCN * wn_iter) + tid_col;
// int local_row = (WM * warp_row + TCM * wm_iter) + tid_row;
// int local_col = (WN * warp_col + TCN * wn_iter) + tid_col;
int local_row = (WM * warp_row);
int local_col = (WN * warp_col);
float *global_offset_C = C +
(BM * threadblock_id_y) * dim_n +
@@ -337,8 +351,7 @@ inline void global_dmem_load(const uint32_t dim_n, const uint32_t dim_k,
const uint32_t local_b_row = tid_in_threadblock / BN;
const uint32_t local_b_col = tid_in_threadblock % BN;
constexpr uint32_t threads_in_warpgroup =
(BM * BN) / ELEM_PER_THREAD / (DOUBLE_BUFFER ? 2 : 1); // FIXME
constexpr uint32_t threads_in_threadblock = (BM * BN) / ELEM_PER_THREAD;
// Data move from GMEM to SMEM
//
@@ -351,7 +364,7 @@ inline void global_dmem_load(const uint32_t dim_n, const uint32_t dim_k,
const uint32_t global_a_row = BM * threadblock_id_y + local_a_row;
// number of rows a full TB can read at a time
constexpr uint32_t row_stride_a = threads_in_warpgroup / BK;
constexpr uint32_t row_stride_a = threads_in_threadblock / BK;
const float *global_a = A + dim_k * global_a_row + (k + local_a_col);
volatile float *local_a_tmp = local_a + BK * local_a_row + local_a_col;
@@ -369,7 +382,7 @@ inline void global_dmem_load(const uint32_t dim_n, const uint32_t dim_k,
}
} else {
if constexpr (!GMEM_COALESCED_A) {
constexpr uint32_t row_stride_as = threads_in_warpgroup / BM;
constexpr uint32_t row_stride_as = threads_in_threadblock / BM;
const uint32_t global_a_row = BM * threadblock_id_y + local_as_col;
const float *global_a = A + dim_k * global_a_row + (k + local_as_row);
// FIXME experimenting with global coalescing
@@ -425,7 +438,7 @@ inline void global_dmem_load(const uint32_t dim_n, const uint32_t dim_k,
local_a_tmp += BM * row_stride_as * 8;
}
} else {
constexpr uint32_t row_stride_a = threads_in_warpgroup / BK;
constexpr uint32_t row_stride_a = threads_in_threadblock / BK;
const uint32_t global_a_row = BM * threadblock_id_y + local_a_row;
const float *global_a = A + dim_k * global_a_row + (k + local_a_col);
// NOTE that SMEM writes are transposed
@@ -478,7 +491,7 @@ inline void global_dmem_load(const uint32_t dim_n, const uint32_t dim_k,
}
}
constexpr uint32_t row_stride_b = threads_in_warpgroup / BN;
constexpr uint32_t row_stride_b = threads_in_threadblock / BN;
const uint32_t global_b_col = BN * threadblock_id_x + local_b_col;
const float *global_b = B + dim_n * (k + local_b_row) + global_b_col;
volatile float *local_b_tmp = local_b + BN * local_b_row + local_b_col;
@@ -524,18 +537,18 @@ inline void global_dmem_load(const uint32_t dim_n, const uint32_t dim_k,
asm volatile ("fsw ft1, %0(%1)" :: "i"(BN * row_stride_b * 1 * sizeof(float)), "r"(local_b_tmp));
asm volatile ("fsw ft2, %0(%1)" :: "i"(BN * row_stride_b * 2 * sizeof(float)), "r"(local_b_tmp));
asm volatile ("fsw ft3, %0(%1)" :: "i"(BN * row_stride_b * 3 * sizeof(float)), "r"(local_b_tmp));
asm volatile ("fsw ft4, %0(%1)" :: "i"(BN * row_stride_b * 4 * sizeof(float)), "r"(local_b_tmp));
asm volatile ("fsw ft5, %0(%1)" :: "i"(BN * row_stride_b * 5 * sizeof(float)), "r"(local_b_tmp));
asm volatile ("fsw ft6, %0(%1)" :: "i"(BN * row_stride_b * 6 * sizeof(float)), "r"(local_b_tmp));
asm volatile ("fsw ft7, %0(%1)" :: "i"(BN * row_stride_b * 7 * sizeof(float)), "r"(local_b_tmp));
local_b_tmp += BN * row_stride_b * 8;
local_b_tmp += BN * row_stride_b * 4;
asm volatile ("fsw ft4, %0(%1)" :: "i"(BN * row_stride_b * 0 * sizeof(float)), "r"(local_b_tmp));
asm volatile ("fsw ft5, %0(%1)" :: "i"(BN * row_stride_b * 1 * sizeof(float)), "r"(local_b_tmp));
asm volatile ("fsw ft6, %0(%1)" :: "i"(BN * row_stride_b * 2 * sizeof(float)), "r"(local_b_tmp));
asm volatile ("fsw ft7, %0(%1)" :: "i"(BN * row_stride_b * 3 * sizeof(float)), "r"(local_b_tmp));
local_b_tmp += BN * row_stride_b * 4;
}
}
inline void thread_block_gemm(kernel_arg_t *__UNIFORM__ arg,
const uint32_t tid_in_threadblock,
const uint32_t threads_per_threadblock,
const uint32_t threadblock_dim_x,
const uint32_t threadblock_dim_y,
/*const uint32_t threadblock_id_x,
const uint32_t threadblock_id_y,*/
@@ -556,17 +569,15 @@ inline void thread_block_gemm(kernel_arg_t *__UNIFORM__ arg,
const uint32_t local_b_row = tid_in_threadblock / BN;
const uint32_t local_b_col = tid_in_threadblock % BN;
const uint32_t threads_per_warpgroup = threads_per_threadblock / (DOUBLE_BUFFER ? 2 : 1);
const uint32_t warpgroup_id = tid_in_threadblock / threads_per_warpgroup;
const uint32_t tid_in_warpgroup = tid_in_threadblock % threads_per_warpgroup; // FIXME
const uint32_t warp_in_warpgroup = tid_in_warpgroup / NUM_LANES;
// FIXME: warp_row / BN should be warp-specialized?
// no double-buffering
const uint32_t threads_per_warpgroup = threads_per_threadblock;
const uint32_t warp_in_warpgroup = threads_per_warpgroup / NUM_LANES;
const uint32_t warp_row = warp_in_warpgroup / (BN / WN);
const uint32_t warp_col = warp_in_warpgroup % (BN / WN);
const uint32_t tid_in_warp = tid_in_threadblock % NUM_LANES;
volatile float *local_a = sharedmem_per_threadblock;
// const size_t local_a_elems = threadblock_dim_x * threadblock_dim_y;
constexpr size_t local_a_elems = (BM * BK);
volatile float *local_b = sharedmem_per_threadblock + local_a_elems;
constexpr size_t local_b_elems = (BK * BN);
@@ -574,125 +585,139 @@ inline void thread_block_gemm(kernel_arg_t *__UNIFORM__ arg,
volatile float *local_a_buf = local_b + local_b_elems;
volatile float *local_b_buf = local_a_buf + local_a_elems;
if (warpgroup_id == 0) {
#pragma GCC unroll 1
for (uint32_t block_m = 0; (block_m * BM) < dim_m; block_m++) {
#pragma GCC unroll 1
for (uint32_t block_n = 0; (block_n * BN) < dim_n; block_n++) {
if constexpr (DOUBLE_BUFFER) {
// initiate software pipeline
global_dmem_load(dim_n, dim_k, 0 /*k*/, A, B, local_a, local_b,
tid_in_warpgroup, block_n, block_m);
constexpr uint32_t skips =
loop_matmul_skips(/*skip_lda=*/0, /*skip_ldb=*/0, /*skip_ldd=*/1,
/*skip_ex=*/1, /*skip_stc=*/1);
threadblock_barrier(0/*threadblock_id_in_cluster*/, threadblock_dim_y);
#if (GEMMINI_DMA == 1)
if (tid_in_threadblock == 0) {
gemmini_extended_config_ex(WEIGHT_STATIONARY, 0, 0, 1, 0, 0);
// gemmini_extended_config_ex(dataflow, act & 3, 0, 1, a_transpose,
// b_transpose);
gemmini_extended3_config_ld(dim_k * sizeof(elem_t), MVIN_SCALE_IDENTITY,
false, 0);
gemmini_extended3_config_ld(dim_n * sizeof(elem_t), MVIN_SCALE_IDENTITY,
false, 1);
gemmini_extended_config_st(dim_n * sizeof(elem_t), 0, MVIN_SCALE_IDENTITY);
gemmini_fence();
}
#endif
#pragma GCC unroll 1
for (uint32_t block_m = 0; (block_m * BM) < dim_m; block_m++) {
#pragma GCC unroll 1
for (uint32_t block_n = 0; (block_n * BN) < dim_n; block_n++) {
// clear out C
initialize_C(0);
initialize_C(1);
// NOTE: this *should* be signed integer to trigger arithmetic
// right-shift
int32_t k_index = 0;
#pragma GCC unroll 1
for (uint32_t block_k = 0; (block_k * BK) < (dim_k); block_k++) {
k_index++;
// producer code: GMEM->SMEM memory movement
// ----------------------------------------------------------------------
#if (GEMMINI_DMA == 1)
if (tid_in_threadblock == 0) {
// configure dma gmem address to load from
// FIXME: block_k is wrong
ROCC_INSTRUCTION_RS1_RS2(
XCUSTOM_ACC,
(uint64_t)(A + block_m * BM * dim_k + block_k * BK),
(uint64_t)(B + block_k * BK * dim_n + block_n * BN),
k_LOOP_WS_CONFIG_ADDRS_AB)
// GEMMINI_CISC(8) does k_LOOP_WS_CONFIG_STRIDES_AB
GEMMINI_CISC_CMD_R((dim_n << 16) | (dim_k << 8) | 8);
// gemmini_fence();
GEMMINI_CISC_CMD_I(13);
// configure loop iteration bounds
// FIXME: shouldn't be necessary
// #define BOUND_INST 0x400040004ULL
// ROCC_INSTRUCTION_RS1_RS2(XCUSTOM_ACC, 0, BOUND_INST,
// k_LOOP_WS_CONFIG_BOUNDS) ROCC_INSTRUCTION_RS1_RS2(XCUSTOM_ACC,
// SPAD_ADDR_Q0, SPAD_ADDR_Q1, k_LOOP_WS_CONFIG_SPAD_AB)
// ROCC_INSTRUCTION_RS1_RS2(
// XCUSTOM_ACC,
// ((uint64_t)(/*a_spad_id:*/ 0) << 18) |
// ((uint64_t)(/*b_spad_id:*/ 0) << 16) |
// ((uint64_t)(/*act:0*/ 0) << 8) | ((/*low_D:*/ 0) << 2) |
// ((/*full_C:*/ 0) << 1) | (/*ex_accumulate:*/ 0),
// ((uint64_t)(/*C_spad_addr:*/ A) << 32) | 0x200U | (skips) |
// ((/*is_resadd*/ 0) << 2) | ((/*B_transpose:*/ 0) << 1) |
// (/*A_transpose:*/ 1),
// k_LOOP_WS)
// gemmini_fence();
// sp_tiled_matmul_full_spad_ws includes CONFIG_BOUNDS
// FIXME: block_k is 0 for two times
// sp_tiled_matmul_full_spad_ws(
// #if 1
// SPAD_ADDR_Q2,
// SPAD_ADDR_Q3,
// #else
// (/*block_k:*/ 0 & 1) ? SPAD_ADDR_Q2 : SPAD_ADDR_Q0,
// (/*block_k:*/ 0 & 1) ? SPAD_ADDR_Q3 : SPAD_ADDR_Q1,
// #endif
// /*spad_D=*/0, /*spad_C=*/SPAD_ADDR_Q1,
// /*I=*/BM / DIM, /*J=*/BN / DIM, /*K=*/BK / DIM, /*pad_I=*/0,
// /*pad_J=*/0, /*pad_K=*/0,
// /*a_transpose=*/1, /*b_transpose=*/0, /*full_C=*/0, /*low_D=*/0,
// /*acc=*/0, /*act=*/NO_ACTIVATION, /*skips=*/skips)
// gemmini_fence();
}
#else
global_dmem_load(dim_n, dim_k, block_k * BK, A, B, local_a, local_b,
tid_in_threadblock, block_n, block_m);
// NOTE: this *should* be signed integer to trigger arithmetic
// right-shift
int32_t k_index = 0;
threadblock_barrier(0 /*threadblock_id_in_cluster*/, threadblock_dim_y);
#endif
// consumer code: SMEM->RF and compute
// ----------------------------------------------------------------------
// @perf: this loop spills to stack a lot because of all the flws in
#pragma GCC unroll 1
for (uint32_t k = 0; k < (8 * dim_k) - BK; k += BK) {
volatile float *local_a_produce;
volatile float *local_b_produce;
if constexpr (DOUBLE_BUFFER) {
const uint32_t mask_odd = (k_index & 1) << 31 >> 31;
const uint32_t mask_even = ((k_index & 1) ^ 1) << 31 >> 31;
// local_a_produce = (k_index % 2) ? local_a : local_a_buf;
// local_b_produce = (k_index % 2) ? local_b : local_b_buf;
local_a_produce = reinterpret_cast<volatile float *>(
(mask_odd & reinterpret_cast<uintmax_t>(local_a)) |
(mask_even & reinterpret_cast<uintmax_t>(local_a_buf)));
local_b_produce = reinterpret_cast<volatile float *>(
(mask_odd & reinterpret_cast<uintmax_t>(local_b)) |
(mask_even & reinterpret_cast<uintmax_t>(local_b_buf)));
} else {
local_a_produce = local_a;
local_b_produce = local_b;
}
k_index++;
global_dmem_load(dim_n, dim_k, k + BK /*runahead*/, A, B,
local_a_produce, local_b_produce, tid_in_warpgroup,
block_n, block_m);
threadblock_barrier(0/*threadblock_id_in_cluster*/, threadblock_dim_y);
}
// sync with final consumer stage in the k-loop
threadblock_barrier(0/*threadblock_id_in_cluster*/, threadblock_dim_y);
}
}
} else {
#pragma GCC unroll 1
for (uint32_t block_m = 0; (block_m * BM) < dim_m; block_m++) {
#pragma GCC unroll 1
for (uint32_t block_n = 0; (block_n * BN) < dim_n; block_n++) {
// clear out C
initialize_C(0);
initialize_C(1);
// sync with initial producer stage in the k-loop
threadblock_barrier(0/*threadblock_id_in_cluster*/, threadblock_dim_y);
// NOTE: this *should* be signed integer to trigger arithmetic
// right-shift
int32_t k_index = 0;
#pragma GCC unroll 1
for (uint32_t k = 0; k < (8 * dim_k); k += BK) {
volatile float *local_a_consume;
volatile float *local_b_consume;
if constexpr (DOUBLE_BUFFER) {
// local_a_consume = (k_index % 2) ? local_a_buf : local_a;
// local_b_consume = (k_index % 2) ? local_b_buf : local_b;
// FIXME: swap multiply with bitshifts
const uint32_t mask_odd = (k_index & 1) << 31 >> 31;
const uint32_t mask_even = ((k_index & 1) ^ 1) << 31 >> 31;
local_a_consume = reinterpret_cast<volatile float *>(
(mask_odd & reinterpret_cast<uintmax_t>(local_a_buf)) |
(mask_even & reinterpret_cast<uintmax_t>(local_a)));
local_b_consume = reinterpret_cast<volatile float *>(
(mask_odd & reinterpret_cast<uintmax_t>(local_b_buf)) |
(mask_even & reinterpret_cast<uintmax_t>(local_b)));
} else {
local_a_consume = local_a;
local_b_consume = local_b;
}
k_index++;
// @perf: this loop spills to stack a lot because of all the flws in
#pragma GCC unroll 1
for (int i = 0; i < BK_LOOP; i++) {
for (int i = 0; i < BK_LOOP; i++) {
#pragma GCC unroll 4
for (uint32_t local_k = 0; local_k < BK; local_k += TCK) {
#pragma GCC unroll 2
for (uint32_t local_k = 0; local_k < BK; local_k += TCK) {
for (int wn_iter = 0; wn_iter < WNITER; wn_iter++) {
// SMEM -> RF
vx_wmma_load_b(local_b, local_k, warp_col, wn_iter, tid_in_warp);
#pragma GCC unroll 2
for (int wn_iter = 0; wn_iter < WNITER; wn_iter++) {
for (int wm_iter = 0; wm_iter < WMITER; wm_iter++) {
// SMEM -> RF
vx_wmma_load_b(local_b_consume, local_k, warp_col, wn_iter,
vx_wmma_load_a(local_a, local_k, warp_row, wm_iter,
tid_in_warp);
#pragma GCC unroll 2
for (int wm_iter = 0; wm_iter < WMITER; wm_iter++) {
// SMEM -> RF
vx_wmma_load_a(local_a_consume, local_k, warp_row, wm_iter,
tid_in_warp);
// perform mma
vx_wmma(wm_iter);
}
// perform mma
vx_wmma(wm_iter);
}
}
}
threadblock_barrier(0/*threadblock_id_in_cluster*/, threadblock_dim_y);
}
#pragma GCC unroll 1
for (int wm_iter = 0; wm_iter < WMITER; wm_iter++) {
#pragma GCC unroll 1
for (int wn_iter = 0; wn_iter < WNITER; wn_iter++) {
if (warpgroup_id == 1) {
write_results(tid_in_warp, warp_col, warp_row, wn_iter, wm_iter,
dim_n, C, block_n, block_m);
}
}
// Call gemmini fence at the end of the loop to overlap dma & wmma.
// Hopefully by this time, dma would have finished so that this is a
// no-op
if (tid_in_threadblock == 0) {
gemmini_fence();
}
threadblock_barrier(0 /*threadblock_id_in_cluster*/, threadblock_dim_y);
}
#pragma GCC unroll 2
for (int wm_iter = 0; wm_iter < WMITER; wm_iter++) {
#pragma GCC unroll 2
for (int wn_iter = 0; wn_iter < WNITER; wn_iter++) {
write_results(tid_in_warp, warp_col, warp_row, wn_iter, wm_iter,
dim_n, C, block_n, block_m);
}
}
}
@@ -703,14 +728,17 @@ 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
const uint32_t threads_per_threadblock = (BM * BN) / (ELEM_PER_THREAD);
// const uint32_t threads_per_threadblock = (BM * BN) / (ELEM_PER_THREAD);
#ifdef RADIANCE
const uint32_t threads_per_threadblock =
CORES_PER_CLUSTER * vx_num_threads() * vx_num_warps();
const uint32_t threadblocks_per_core = CORES_PER_CLUSTER * vx_num_threads() *
vx_num_warps() /
threads_per_threadblock;
#else
const uint32_t threadblocks_per_core =
vx_num_threads() * vx_num_warps() / threads_per_threadblock;
const uint32_t threads_per_threadblock = vx_num_threads() * vx_num_warps();
const uint32_t threadblocks_per_core =
vx_num_threads() * vx_num_warps() / threads_per_threadblock;
#endif
const uint32_t threadblock_dim_x = vx_num_threads();
const uint32_t threadblock_dim_y = vx_num_warps() / threadblocks_per_core;
@@ -731,7 +759,7 @@ void kernel_body(int task_id, kernel_arg_t *__UNIFORM__ arg) {
const int warp_id = vx_warp_id();
thread_block_gemm(arg, tid_in_threadblock, threads_per_threadblock,
threadblock_dim_x, threadblock_dim_y, /*threadblock_id_x,
threadblock_dim_y, /*threadblock_id_x,
threadblock_id_y,*/ /*threadblock_id_in_cluster, */
sharedmem_per_threadblock);
}