Skip to content

metal: template for mat-vec multiplication kernels #2891

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 6 commits into
base: master
Choose a base branch
from
Open

metal: template for mat-vec multiplication kernels #2891

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
bca5d0c
metal: template for mat-vec multiplication kernels
lshzh-ww Aug 30, 2023
804c78d
metal : minor indentations
ggerganov Aug 30, 2023
aa4b7d2
metal: improvement for Q4_K driver
lshzh-ww Aug 31, 2023
35cd10c
metal: yet another MUL mat-vec template
lshzh-ww Aug 31, 2023
048831e
Merge branch 'master' into HEAD
ggerganov Sep 1, 2023
b46ae7b
Merge branch 'master' into HEAD
ggerganov Sep 1, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
202 changes: 67 additions & 135 deletions ggml-metal.m
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -76,14 +76,15 @@
GGML_METAL_DECL_KERNEL(rms_norm);
GGML_METAL_DECL_KERNEL(norm);
GGML_METAL_DECL_KERNEL(mul_mat_f16_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q4_0_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q4_1_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q8_0_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q2_K_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q3_K_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q4_K_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q5_K_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q6_K_f32);
GGML_METAL_DECL_KERNEL(mul_mv_f16_f32);
GGML_METAL_DECL_KERNEL(mul_mv_q4_0_f32);
GGML_METAL_DECL_KERNEL(mul_mv_q4_1_f32);
GGML_METAL_DECL_KERNEL(mul_mv_q8_0_f32);
GGML_METAL_DECL_KERNEL(mul_mv_q2_K_f32);
GGML_METAL_DECL_KERNEL(mul_mv_q3_K_f32);
GGML_METAL_DECL_KERNEL(mul_mv_q4_K_f32);
GGML_METAL_DECL_KERNEL(mul_mv_q5_K_f32);
GGML_METAL_DECL_KERNEL(mul_mv_q6_K_f32);
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This rename is OK

GGML_METAL_DECL_KERNEL(mul_mm_f16_f32);
GGML_METAL_DECL_KERNEL(mul_mm_q4_0_f32);
GGML_METAL_DECL_KERNEL(mul_mm_q4_1_f32);
Expand Down Expand Up @@ -205,14 +206,15 @@ @implementation GGMLMetalClass
GGML_METAL_ADD_KERNEL(rms_norm);
GGML_METAL_ADD_KERNEL(norm);
GGML_METAL_ADD_KERNEL(mul_mat_f16_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q4_0_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q4_1_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q8_0_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q2_K_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q3_K_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q4_K_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q5_K_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q6_K_f32);
GGML_METAL_ADD_KERNEL(mul_mv_f16_f32);
GGML_METAL_ADD_KERNEL(mul_mv_q4_0_f32);
GGML_METAL_ADD_KERNEL(mul_mv_q4_1_f32);
GGML_METAL_ADD_KERNEL(mul_mv_q8_0_f32);
GGML_METAL_ADD_KERNEL(mul_mv_q2_K_f32);
GGML_METAL_ADD_KERNEL(mul_mv_q3_K_f32);
GGML_METAL_ADD_KERNEL(mul_mv_q4_K_f32);
GGML_METAL_ADD_KERNEL(mul_mv_q5_K_f32);
GGML_METAL_ADD_KERNEL(mul_mv_q6_K_f32);
GGML_METAL_ADD_KERNEL(mul_mm_f16_f32);
GGML_METAL_ADD_KERNEL(mul_mm_q4_0_f32);
GGML_METAL_ADD_KERNEL(mul_mm_q8_0_f32);
Expand Down Expand Up @@ -270,14 +272,15 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
GGML_METAL_DEL_KERNEL(rms_norm);
GGML_METAL_DEL_KERNEL(norm);
GGML_METAL_DEL_KERNEL(mul_mat_f16_f32);
GGML_METAL_DEL_KERNEL(mul_mat_q4_0_f32);
GGML_METAL_DEL_KERNEL(mul_mat_q4_1_f32);
GGML_METAL_DEL_KERNEL(mul_mat_q8_0_f32);
GGML_METAL_DEL_KERNEL(mul_mat_q2_K_f32);
GGML_METAL_DEL_KERNEL(mul_mat_q3_K_f32);
GGML_METAL_DEL_KERNEL(mul_mat_q4_K_f32);
GGML_METAL_DEL_KERNEL(mul_mat_q5_K_f32);
GGML_METAL_DEL_KERNEL(mul_mat_q6_K_f32);
GGML_METAL_DEL_KERNEL(mul_mv_f16_f32);
GGML_METAL_DEL_KERNEL(mul_mv_q4_0_f32);
GGML_METAL_DEL_KERNEL(mul_mv_q4_1_f32);
GGML_METAL_DEL_KERNEL(mul_mv_q8_0_f32);
GGML_METAL_DEL_KERNEL(mul_mv_q2_K_f32);
GGML_METAL_DEL_KERNEL(mul_mv_q3_K_f32);
GGML_METAL_DEL_KERNEL(mul_mv_q4_K_f32);
GGML_METAL_DEL_KERNEL(mul_mv_q5_K_f32);
GGML_METAL_DEL_KERNEL(mul_mv_q6_K_f32);
GGML_METAL_DEL_KERNEL(mul_mm_f16_f32);
GGML_METAL_DEL_KERNEL(mul_mm_q4_0_f32);
GGML_METAL_DEL_KERNEL(mul_mm_q8_0_f32);
Expand Down Expand Up @@ -844,97 +847,45 @@ void ggml_metal_graph_compute(
[encoder setBytes:&gqa length:sizeof(gqa) atIndex:10];
[encoder setThreadgroupMemoryLength:8192 atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake( (ne11+31)/32, (ne01+63) / 64, ne12) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
} else {
int nth0 = 32;
int nth1 = 1;

} else if (ggml_is_contiguous(src0) &&
ggml_is_contiguous(src1) &&
src1t == GGML_TYPE_F32 &&
ne00%32 == 0) {
// use custom matrix x vector kernel
switch (src0t) {
case GGML_TYPE_F16:
{
nth0 = 32;
nth1 = 1;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32];
} break;
case GGML_TYPE_Q4_0:
{
GGML_ASSERT(ne02 == 1);
GGML_ASSERT(ne12 == 1);

nth0 = 8;
nth1 = 8;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_0_f32];
} break;
case GGML_TYPE_Q4_1:
{
GGML_ASSERT(ne02 == 1);
GGML_ASSERT(ne12 == 1);

nth0 = 8;
nth1 = 8;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_1_f32];
} break;
case GGML_TYPE_Q8_0:
{
GGML_ASSERT(ne02 == 1);
GGML_ASSERT(ne12 == 1);

nth0 = 8;
nth1 = 8;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_q8_0_f32];
} break;
case GGML_TYPE_Q2_K:
{
GGML_ASSERT(ne02 == 1);
GGML_ASSERT(ne12 == 1);

nth0 = 2;
nth1 = 32;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_q2_K_f32];
} break;
case GGML_TYPE_Q3_K:
{
GGML_ASSERT(ne02 == 1);
GGML_ASSERT(ne12 == 1);

nth0 = 2;
nth1 = 32;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_q3_K_f32];
} break;
case GGML_TYPE_Q4_K:
{
GGML_ASSERT(ne02 == 1);
GGML_ASSERT(ne12 == 1);

nth0 = 2;
nth1 = 32;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_K_f32];
} break;
case GGML_TYPE_Q5_K:
{
GGML_ASSERT(ne02 == 1);
GGML_ASSERT(ne12 == 1);

nth0 = 2;
nth1 = 32;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_q5_K_f32];
} break;
case GGML_TYPE_Q6_K:
{
GGML_ASSERT(ne02 == 1);
GGML_ASSERT(ne12 == 1);

nth0 = 2;
nth1 = 32;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_q6_K_f32];
} break;
default:
{
metal_printf("Asserting on type %d\n",(int)src0t);
GGML_ASSERT(false && "not implemented");
}
};

switch (src0->type) {
case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_mul_mv_f16_f32]; break;
case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_mul_mv_q4_0_f32]; break;
case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_mul_mv_q4_1_f32]; break;
case GGML_TYPE_Q8_0: [encoder setComputePipelineState:ctx->pipeline_mul_mv_q8_0_f32]; break;
case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_mul_mv_q2_K_f32]; break;
case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_mul_mv_q3_K_f32]; break;
case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_mul_mv_q4_K_f32]; break;
case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_mul_mv_q5_K_f32]; break;
case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_mul_mv_q6_K_f32]; break;
default: GGML_ASSERT(false && "MUL MAT-VEC not implemented");
}
int buffer_size_aligned = (512 / ggml_blck_size(src0t) * ggml_element_size(src0) + 31) / 32 * 32;
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
[encoder setBytes:&ne10 length:sizeof(ne10) atIndex:6];
[encoder setBytes:&ne12 length:sizeof(ne12) atIndex:7];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:8];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:9];
[encoder setBytes:&gqa length:sizeof(gqa) atIndex:10];
// only for k-quants we use threadgroup memory
if (ggml_blck_size(src0t) >= 64){
[encoder setThreadgroupMemoryLength:8 * buffer_size_aligned atIndex:0];
}
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(64, 1, 1)];
} else {
switch (src0->type) {
case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32]; break;
default: GGML_ASSERT(false && " not implemented");
}
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
Expand All @@ -953,27 +904,8 @@ void ggml_metal_graph_compute(
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:15];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:16];
[encoder setBytes:&gqa length:sizeof(gqa) atIndex:17];

if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 || src0t == GGML_TYPE_Q8_0 ||
src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_Q4_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_Q3_K) {
#ifdef GGML_QKK_64
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
#else
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
#endif
}
else if (src0t == GGML_TYPE_Q5_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_Q6_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
} else {
[encoder setThreadgroupMemoryLength:nth0*sizeof(float) atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
[encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(32, 1, 1)];
}
} break;
case GGML_OP_GET_ROWS:
Expand Down
Loading