cuBLAS: use host pinned memory and dequantize while copying (#1207)
* cuBLAS: dequantize simultaneously while copying memory * cuBLAS: use host pinned memory * cuBLAS: improve ggml_compute_forward_mul_mat_f16_f32 with pinned memory * cuBLAS: also pin kv cache * fix rebase
This commit is contained in:
slaren
GitHub
parent
b1ee8f59b4
commit
7fc50c051a
6 changed files with 110 additions and 54 deletions
70
ggml.c
70
ggml.c
|
|
@ -8033,7 +8033,7 @@ static void ggml_compute_forward_mul_mat_f32(
|
|||
#if defined(GGML_USE_CUBLAS)
|
||||
const float alpha = 1.0f;
|
||||
const float beta = 0.0f;
|
||||
const int x_ne = ne01 * ne10;
|
||||
const int x_ne = ne01 * ne00;
|
||||
const int y_ne = ne11 * ne10;
|
||||
const int d_ne = ne11 * ne01;
|
||||
|
||||
|
|
@ -8235,25 +8235,27 @@ static void ggml_compute_forward_mul_mat_f16_f32(
|
|||
}
|
||||
|
||||
#if defined(GGML_USE_CUBLAS)
|
||||
ggml_fp16_t * const wdata = params->wdata;
|
||||
|
||||
const float alpha = 1.0f;
|
||||
const float beta = 0.0f;
|
||||
const int x_ne = ne01 * ne10;
|
||||
const int x_ne = ne01 * ne00;
|
||||
const int y_ne = ne11 * ne10;
|
||||
const int d_ne = ne11 * ne01;
|
||||
|
||||
size_t x_size, y_size, d_size;
|
||||
float *d_X = ggml_cuda_pool_malloc(sizeof(float) * x_ne, &x_size);
|
||||
float *d_Y = ggml_cuda_pool_malloc(sizeof(float) * y_ne, &y_size);
|
||||
float *d_D = ggml_cuda_pool_malloc(sizeof(float) * d_ne, &d_size);
|
||||
ggml_fp16_t * d_X = ggml_cuda_pool_malloc(sizeof(float) * x_ne, &x_size);
|
||||
ggml_fp16_t * d_Y = ggml_cuda_pool_malloc(sizeof(float) * y_ne, &y_size);
|
||||
float * d_D = ggml_cuda_pool_malloc(sizeof(float) * d_ne, &d_size);
|
||||
#else
|
||||
float * const wdata = params->wdata;
|
||||
#endif
|
||||
for (int64_t i03 = 0; i03 < ne03; i03++) {
|
||||
for (int64_t i02 = 0; i02 < ne02; i02++) {
|
||||
#if defined(GGML_USE_CUBLAS)
|
||||
// copy src0 while converting src1
|
||||
CUDA_CHECK(ggml_cuda_h2d_tensor_2d(d_X, src0, i03, i02, g_cudaStream));
|
||||
|
||||
// with cuBlAS, instead of converting src0 to fp32, we convert src1 to fp16
|
||||
ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + (ne11 * ne10) * (i03 * ne02 + i02);
|
||||
{
|
||||
size_t id = 0;
|
||||
for (int64_t i01 = 0; i01 < ne11; ++i01) {
|
||||
|
|
@ -8275,11 +8277,9 @@ static void ggml_compute_forward_mul_mat_f16_f32(
|
|||
|
||||
#if defined(GGML_USE_CUBLAS)
|
||||
const ggml_fp16_t * y = (ggml_fp16_t *) wdata;
|
||||
|
||||
float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
|
||||
|
||||
// copy data to device
|
||||
CUDA_CHECK(ggml_cuda_h2d_tensor_2d(d_X, src0, i03, i02, g_cudaStream));
|
||||
CUDA_CHECK(cudaMemcpyAsync(d_Y, y, sizeof(ggml_fp16_t) * y_ne, cudaMemcpyHostToDevice, g_cudaStream));
|
||||
|
||||
// compute
|
||||
|
|
@ -8498,39 +8498,19 @@ static void ggml_compute_forward_mul_mat_q_f32(
|
|||
#if defined(GGML_USE_CUBLAS)
|
||||
const float alpha = 1.0f;
|
||||
const float beta = 0.0f;
|
||||
const int x_ne = ne01 * ne10;
|
||||
const int x_ne = ne01 * ne00;
|
||||
const int y_ne = ne11 * ne10;
|
||||
const int d_ne = ne11 * ne01;
|
||||
|
||||
size_t x_size, y_size, d_size, q_size;
|
||||
float *d_X = ggml_cuda_pool_malloc(sizeof(float) * x_ne, &x_size);
|
||||
float *d_Y = ggml_cuda_pool_malloc(sizeof(float) * y_ne, &y_size);
|
||||
float *d_D = ggml_cuda_pool_malloc(sizeof(float) * d_ne, &d_size);
|
||||
float *d_Q = ggml_cuda_pool_malloc(GGML_TYPE_SIZE[type] * x_ne / GGML_BLCK_SIZE[type], &q_size);
|
||||
float * d_X = ggml_cuda_pool_malloc(sizeof(float) * x_ne, &x_size);
|
||||
float * d_Y = ggml_cuda_pool_malloc(sizeof(float) * y_ne, &y_size);
|
||||
float * d_D = ggml_cuda_pool_malloc(sizeof(float) * d_ne, &d_size);
|
||||
void * d_Q = ggml_cuda_pool_malloc(GGML_TYPE_SIZE[type] * x_ne / GGML_BLCK_SIZE[type], &q_size);
|
||||
|
||||
void (*dequantize_row_q_cuda)(const void * x, float * y, int k, cudaStream_t stream) = NULL;
|
||||
if (type == GGML_TYPE_Q4_0) {
|
||||
dequantize_row_q_cuda = dequantize_row_q4_0_cuda;
|
||||
}
|
||||
else if (type == GGML_TYPE_Q4_1) {
|
||||
dequantize_row_q_cuda = dequantize_row_q4_1_cuda;
|
||||
}
|
||||
else if (type == GGML_TYPE_Q4_2) {
|
||||
dequantize_row_q_cuda = dequantize_row_q4_2_cuda;
|
||||
}
|
||||
else if (type == GGML_TYPE_Q5_0) {
|
||||
dequantize_row_q_cuda = dequantize_row_q5_0_cuda;
|
||||
}
|
||||
else if (type == GGML_TYPE_Q5_1) {
|
||||
dequantize_row_q_cuda = dequantize_row_q5_1_cuda;
|
||||
}
|
||||
else if (type == GGML_TYPE_Q8_0) {
|
||||
dequantize_row_q_cuda = dequantize_row_q8_0_cuda;
|
||||
}
|
||||
else {
|
||||
GGML_ASSERT(false);
|
||||
}
|
||||
#elif !defined(GGML_USE_CLBLAST)
|
||||
const dequantize_row_q_cuda_t dequantize_row_q_cuda = ggml_get_dequantize_row_q_cuda(type);
|
||||
GGML_ASSERT(dequantize_row_q_cuda != NULL);
|
||||
#else
|
||||
float * const wdata = params->wdata;
|
||||
dequantize_row_q_t const dequantize_row_q = quantize_fns[type].dequantize_row_q;
|
||||
#endif
|
||||
|
|
@ -8543,10 +8523,11 @@ static void ggml_compute_forward_mul_mat_q_f32(
|
|||
|
||||
#if defined(GGML_USE_CUBLAS)
|
||||
// copy and dequantize on device
|
||||
CUDA_CHECK(ggml_cuda_h2d_tensor_2d(d_Q, src0, i03, i02, g_cudaStream));
|
||||
CUDA_CHECK(ggml_cuda_h2d_tensor_2d(d_Q, src0, i03, i02, g_cudaStream2));
|
||||
|
||||
dequantize_row_q_cuda(d_Q, d_X, ne01 * ne00, g_cudaStream);
|
||||
dequantize_row_q_cuda(d_Q, d_X, x_ne, g_cudaStream2);
|
||||
CUDA_CHECK(cudaGetLastError());
|
||||
CUDA_CHECK(cudaEventRecord(g_cudaEvent, g_cudaStream2));
|
||||
#elif defined(GGML_USE_CLBLAST)
|
||||
const void* x = (char *) src0->data + i03*nb03 + i02*nb02;
|
||||
#else
|
||||
|
|
@ -8560,11 +8541,13 @@ static void ggml_compute_forward_mul_mat_q_f32(
|
|||
const float * x = wdata;
|
||||
#endif
|
||||
|
||||
|
||||
#if defined(GGML_USE_CUBLAS)
|
||||
// copy data to device
|
||||
CUDA_CHECK(ggml_cuda_h2d_tensor_2d(d_Y, src1, i03, i02, g_cudaStream));
|
||||
|
||||
// wait for dequantization
|
||||
CUDA_CHECK(cudaStreamWaitEvent(g_cudaStream, g_cudaEvent, 0));
|
||||
|
||||
// compute
|
||||
CUBLAS_CHECK(
|
||||
cublasSgemm(g_cublasH, CUBLAS_OP_T, CUBLAS_OP_N,
|
||||
|
|
@ -11588,7 +11571,7 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
|
|||
if (ggml_compute_forward_mul_mat_use_blas(node->src0, node->src1, node)) {
|
||||
node->n_tasks = 1; // TODO: this actually is doing nothing
|
||||
// the threads are still spinning
|
||||
cur = GGML_TYPE_SIZE[GGML_TYPE_F32]*(node->src0->ne[0]*node->src0->ne[1]);
|
||||
cur = GGML_TYPE_SIZE[GGML_TYPE_F32]*MAX(ggml_nelements(node->src1), ggml_nelements(node->src0));
|
||||
//printf("src0: ne0 = %d, ne1 = %d, ne = %d\n", node->src0->ne[0], node->src0->ne[1], node->src0->ne[0]*node->src0->ne[1]);
|
||||
//printf("src1: ne0 = %d, ne1 = %d, ne = %d\n", node->src1->ne[0], node->src1->ne[1], node->src1->ne[0]*node->src1->ne[1]);
|
||||
//printf("cur = %zu\n", cur);
|
||||
|
|
@ -11600,6 +11583,11 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
|
|||
#endif
|
||||
} else if (node->src0->type == GGML_TYPE_F32 && node->src1->type == GGML_TYPE_F32) {
|
||||
cur = 0;
|
||||
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS)
|
||||
if (ggml_compute_forward_mul_mat_use_blas(node->src0, node->src1, node)) {
|
||||
node->n_tasks = 1;
|
||||
}
|
||||
#endif
|
||||
} else if (ggml_is_quantized(node->src0->type) && node->src1->type == GGML_TYPE_F32) {
|
||||
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
|
||||
if (ggml_compute_forward_mul_mat_use_blas(node->src0, node->src1, node)) {
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue