metal : add special-case mat-vec mul for ne00 == 4 (#14385)

ggml-ci
This commit is contained in:
Georgi Gerganov
2025-06-26 15:51:19 +03:00
committed by GitHub
parent 5783ae4359
commit e8215dbb96
3 changed files with 121 additions and 32 deletions

View File

@@ -211,11 +211,14 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_RWKV_WKV6_F32,
GGML_METAL_KERNEL_TYPE_RWKV_WKV7_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32_C4,
GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_C4,
GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW,
GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4,
GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16,
GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_C4,
GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_1ROW,
GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_L4,
GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_BF16,
@@ -1175,11 +1178,14 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RWKV_WKV6_F32, rwkv_wkv6_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RWKV_WKV7_F32, rwkv_wkv7_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32, mul_mv_f32_f32, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32_C4, mul_mv_f32_f32_c4, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32, mul_mv_bf16_f32, has_simdgroup_reduction && use_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_C4, mul_mv_bf16_f32_c4, use_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_1ROW, mul_mv_bf16_f32_1row, has_simdgroup_reduction && use_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_L4, mul_mv_bf16_f32_l4, has_simdgroup_reduction && use_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_BF16, mul_mv_bf16_bf16, has_simdgroup_reduction && use_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32, mul_mv_f16_f32, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_C4, mul_mv_f16_f32_c4, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW, mul_mv_f16_f32_1row, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4, mul_mv_f16_f32_l4, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16, mul_mv_f16_f16, has_simdgroup_reduction);
@@ -3111,14 +3117,23 @@ static bool ggml_metal_encode_node(
nsg = 1;
nr0 = 1;
nr1 = 4;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32].pipeline;
if (ne00 == 4) {
nr0 = 32;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32_C4].pipeline;
} else {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32].pipeline;
}
} break;
case GGML_TYPE_F16:
{
nsg = 1;
nr0 = 1;
if (src1t == GGML_TYPE_F32) {
if (ne11 * ne12 < 4) {
if (ne00 == 4) {
nr0 = 32;
nr1 = 4;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_C4].pipeline;
} else if (ne11 * ne12 < 4) {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW].pipeline;
} else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4].pipeline;
@@ -3137,7 +3152,11 @@ static bool ggml_metal_encode_node(
nsg = 1;
nr0 = 1;
if (src1t == GGML_TYPE_F32) {
if (ne11 * ne12 < 4) {
if (ne00 == 4) {
nr0 = 32;
nr1 = 4;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_C4].pipeline;
} else if (ne11 * ne12 < 4) {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_1ROW].pipeline;
} else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_L4].pipeline;