vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

move BLAS to a separate backend (#6210)

Browse source 
* move BLAS to a separate backend

* rename GGML_USE_OPENBLAS to GGML_USE_BLAS

* alloc : reuse same buffer when the same buffer type if used multiple times

* set number of threads automatically for openblas and blis

* sched : print assignments when GGML_SCHED_DEBUG env variable is set

* sched : allow ops with weights on an incompatible buffer type

This will cause the weight to be copied to a backend that supports the
op, which is very costly. The weight should have been stored in a buffer
of a backend that can run the op, but llama.cpp cannot do this
automatically at the moment.

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
slaren 2024-06-13 03:11:35 +02:00 • committed by GitHub
parent 1c641e6aac
commit f578b86b21
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
17 changed files with 821 additions and 379 deletions
Download patch file Download diff file Expand all files Collapse all files

205
ggml.c
View file

@ -297,12 +297,6 @@ inline static void * ggml_calloc(size_t num, size_t size) {
#if defined(GGML_USE_ACCELERATE)
#include <Accelerate/Accelerate.h>
#elif defined(GGML_USE_OPENBLAS)
#if defined(GGML_BLAS_USE_MKL)
#include <mkl.h>
#else
#include <cblas.h>
#endif
#endif
// floating point type used to accumulate sums
@ -12179,39 +12173,6 @@ static void ggml_compute_forward_group_norm(
// ggml_compute_forward_mul_mat
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
// helper function to determine if it is better to use BLAS or not
// for large matrices, BLAS is faster
static bool ggml_compute_forward_mul_mat_use_blas(struct ggml_tensor * dst) {
const struct ggml_tensor * src0 = dst->src[0];
const struct ggml_tensor * src1 = dst->src[1];
//const int64_t ne00 = src0->ne[0];
//const int64_t ne01 = src0->ne[1];
const int64_t ne10 = src1->ne[0];
const int64_t ne0 = dst->ne[0];
const int64_t ne1 = dst->ne[1];
// NOTE: with GGML_OP_MUL_MAT_ID we don't want to go through the BLAS branch because it will dequantize (to_float)
// all the experts for each batch element and the processing would become incredibly slow
// TODO: find the optimal values for these
if (dst->op != GGML_OP_MUL_MAT_ID &&
ggml_is_contiguous(src0) &&
ggml_is_contiguous(src1) &&
//src0->type == GGML_TYPE_F32 &&
src1->type == GGML_TYPE_F32 &&
(ne0 >= 32 && ne1 >= 32 && ne10 >= 32)) {
/*printf("BLAS: %d %d %d %d %d\n", ne0, ne1, ne10, ne00, ne01);*/
return true;
}
return false;
}
#endif
static void ggml_compute_forward_mul_mat_one_chunk(
const struct ggml_compute_params * params,
struct ggml_tensor * dst,
@ -12349,73 +12310,6 @@ static void ggml_compute_forward_mul_mat(
// nb01 >= nb00 - src0 is not transposed
// compute by src0 rows
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
if (ggml_compute_forward_mul_mat_use_blas(dst)) {
const int64_t ne_plane = ne01*ne00;
const size_t desired_wsize = ne13*ne12*ne_plane*sizeof(float);
UNUSED(desired_wsize);
if (params->type == GGML_TASK_TYPE_INIT) {
if (type != GGML_TYPE_F32) {
assert(params->wsize >= desired_wsize);
// parallelize by src0 rows
for (int64_t i13 = 0; i13 < ne13; i13++) {
for (int64_t i12 = 0; i12 < ne12; i12++) {
// broadcast src0 into src1 across 2nd,3rd dimension
const int64_t i03 = i13/r3;
const int64_t i02 = i12/r2;
const void * x = (char *) src0->data + i02*nb02 + i03*nb03;
float * const wdata = (float *) params->wdata + i13*ne12*ne_plane + i12*ne_plane;
ggml_to_float_t const to_float = type_traits[type].to_float;
for (int64_t i01 = ith; i01 < ne01; i01 += nth) {
to_float((const char *) x + i01*nb01, wdata + i01*ne00, ne00);
}
}
}
}
return;
}
if (params->type == GGML_TASK_TYPE_FINALIZE) {
return;
}
// perform sgemm, parallelization controlled by blas lib
if (ith != 0) {
return;
}
//const int64_t tgemm0 = ggml_perf_time_us();
for (int64_t i13 = 0; i13 < ne13; i13++) {
for (int64_t i12 = 0; i12 < ne12; i12++) {
const int64_t i03 = i13/r3;
const int64_t i02 = i12/r2;
const void * x = (char *) src0->data + i02*nb02 + i03*nb03;
const float * y = (float *) ((char *) src1->data + i12*nb12 + i13*nb13);
float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
if (type != GGML_TYPE_F32) {
x = (float *) params->wdata + i13*ne12*ne_plane + i12*ne_plane;
}
cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
ne1, ne01, ne10,
1.0f, y, ne10,
x, ne00,
0.0f, d, ne01);
}
}
//printf("cblas_sgemm = %.3f ms, %lld flops\n", (ggml_perf_time_us() - tgemm0)/1000.0, ne13*ne12*ne1*ne01*ne10*2);
//printf("CBLAS = %f ms, %d x %d x %d x %d\n", (ggml_perf_time_us() - t0)/1000.0, ne0, ne1, ne2, ne3);
return;
}
#endif
#if GGML_USE_LLAMAFILE
const bool src1_cont = ggml_is_contiguous(src1);
@ -12796,19 +12690,7 @@ static void ggml_compute_forward_out_prod_f32(
// nb01 >= nb00 - src0 is not transposed
// compute by src0 rows
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
bool use_blas = ggml_is_matrix(src0) &&
ggml_is_matrix(src1) &&
ggml_is_contiguous(src0) &&
(ggml_is_contiguous(src1) || ggml_is_transposed(src1));
#endif
if (params->type == GGML_TASK_TYPE_INIT) {
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) // gemm beta will zero dst
if (use_blas) {
return;
}
#endif
if (ith != 0) {
return;
}
@ -12820,50 +12702,6 @@ static void ggml_compute_forward_out_prod_f32(
return;
}
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
if (use_blas) {
if (params->ith != 0) { // All threads other than the first do no work.
return;
}
// Arguments to ggml_compute_forward_out_prod (expressed as major,minor)
// src0: (k,n)
// src1: (k,m)
// dst: (m,n)
//
// Arguments to sgemm (see https://github.com/Reference-LAPACK/lapack/blob/master/BLAS/SRC/sgemm.f)
// Also expressed as (major,minor)
// a: (m,k): so src1 transposed
// b: (k,n): so src0
// c: (m,n)
//
// However, if ggml_is_transposed(src1) is true, then
// src1->data already contains a transposed version, so sgemm mustn't
// transpose it further.
int n = src0->ne[0];
int k = src0->ne[1];
int m = src1->ne[0];
int transposeA, lda;
if (!ggml_is_transposed(src1)) {
transposeA = CblasTrans;
lda = m;
} else {
transposeA = CblasNoTrans;
lda = k;
}
float * a = (float *) ((char *) src1->data);
float * b = (float *) ((char *) src0->data);
float * c = (float *) ((char *) dst->data);
cblas_sgemm(CblasRowMajor, transposeA, CblasNoTrans, m, n, k, 1.0, a, lda, b, n, 0.0, c, n);
return;
}
#endif
// dst[:,:,:,:] = 0
// for i2,i3:
// for i1:
@ -12993,8 +12831,6 @@ static void ggml_compute_forward_out_prod_q_f32(
// nb01 >= nb00 - src0 is not transposed
// compute by src0 rows
// TODO: #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
if (params->type == GGML_TASK_TYPE_INIT) {
if (ith != 0) {
return;
@ -13391,6 +13227,8 @@ static void ggml_compute_forward_get_rows_q(
const int64_t i10 = (i - i12*ne11*ne10 - i11*ne10);
const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
assert(i01 >= 0 && i01 < ne01);
dequantize_row_q(
(const void *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03),
(float *) ((char *) dst->data + i10*nb1 + i11*nb2 + i12*nb3), nc);
@ -13434,6 +13272,8 @@ static void ggml_compute_forward_get_rows_f16(
const int64_t i10 = (i - i12*ne11*ne10 - i11*ne10);
const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
assert(i01 >= 0 && i01 < ne01);
ggml_fp16_to_fp32_row(
(const void *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03),
(float *) ((char *) dst->data + i10*nb1 + i11*nb2 + i12*nb3), nc);
@ -13477,7 +13317,9 @@ static void ggml_compute_forward_get_rows_bf16(
const int64_t i10 = (i - i12*ne11*ne10 - i11*ne10);
const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
ggml_bf16_to_fp32_row(
assert(i01 >= 0 && i01 < ne01);
ggml_bf16_to_fp32_row(
(const void *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03),
(float *) ((char *) dst->data + i10*nb1 + i11*nb2 + i12*nb3), nc);
}
@ -13520,6 +13362,8 @@ static void ggml_compute_forward_get_rows_f32(
const int64_t i10 = (i - i12*ne11*ne10 - i11*ne10);
const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
assert(i01 >= 0 && i01 < ne01);
ggml_vec_cpy_f32(nc,
(float *) ((char *) dst->data + i10*nb1 + i11*nb2 + i12*nb3),
(float *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03));
@ -18893,6 +18737,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads, int n_cur_
switch (node->op) {
case GGML_OP_CPY:
case GGML_OP_DUP:
case GGML_OP_CONT:
case GGML_OP_ADD:
case GGML_OP_ADD1:
case GGML_OP_ACC:
@ -18977,7 +18822,6 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads, int n_cur_
} break;
case GGML_OP_SCALE:
case GGML_OP_SET:
case GGML_OP_CONT:
case GGML_OP_RESHAPE:
case GGML_OP_VIEW:
case GGML_OP_PERMUTE:
@ -19137,8 +18981,11 @@ static void ggml_graph_compute_thread_sync_node(int * node_n, struct ggml_comput
sched_yield();
}
* node_n = atomic_load(&state->shared->node_n);
if (* node_n != last_node_n) break;
*node_n = atomic_load(&state->shared->node_n);
if (*node_n != last_node_n) {
break;
}
#if defined(__SSE3__)
// Tell the processor we're spinning. It's a processor hint for spinlocks.
_mm_pause();
@ -19148,15 +18995,18 @@ static void ggml_graph_compute_thread_sync_node(int * node_n, struct ggml_comput
static void ggml_graph_compute_thread_sync_task(int * task_phase, struct ggml_compute_state * state, const bool do_yield) {
// wait for other threads to finish
const int last_task_phase = * task_phase;
const int last_task_phase = *task_phase;
while (true) {
if (do_yield) {
sched_yield();
}
* task_phase = atomic_load(&state->shared->node_task);
if (* task_phase != last_task_phase) break;
*task_phase = atomic_load(&state->shared->node_task);
if (*task_phase != last_task_phase) {
break;
}
#if defined(__SSE3__)
// Tell the processor we're spinning. It's a processor hint for spinlocks.
_mm_pause();
@ -19356,17 +19206,6 @@ struct ggml_cplan ggml_graph_plan(const struct ggml_cgraph * cgraph, int n_threa
{
const enum ggml_type vec_dot_type = type_traits[node->src[0]->type].vec_dot_type;
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
if (ggml_compute_forward_mul_mat_use_blas(node)) {
if (node->src[0]->type != GGML_TYPE_F32) {
// here we need memory for fully dequantized matrix from src0
// take into account that src0 can be broadcasted into src1[2,3]
cur = ggml_type_size(GGML_TYPE_F32)
* node->src[0]->ne[0]*node->src[0]->ne[1]
* node->src[1]->ne[2]*node->src[1]->ne[3];
}
} else
#endif
if (node->src[1]->type != vec_dot_type) {
cur = ggml_row_size(vec_dot_type, ggml_nelements(node->src[1]));
}
@ -22664,7 +22503,7 @@ int ggml_cpu_has_wasm_simd(void) {
}
int ggml_cpu_has_blas(void) {
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUDA) || defined(GGML_USE_VULKAN) || defined(GGML_USE_SYCL)
#if defined(GGML_USE_BLAS) || defined(GGML_USE_CUDA) || defined(GGML_USE_VULKAN) || defined(GGML_USE_SYCL)
return 1;
#else
return 0;