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

Added kernel_mul_mv_f16_f32_l4_large which performs 32x more ops

Browse source 
Most of the time, kernel_mul_mv_f16_f32_l4 is called to perform 4 FP ops per thread. Added kernel_mul_mv_f16_f32_l4_large which performs 128 FP ops per thread, when there are 32x less threads.
This commit is contained in:
Alexander Komarov 2024-05-24 11:50:23 -07:00 • committed by GitHub
parent d041d2ceaa
commit cd2322c996
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 12 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

14
ggml-metal.m
View file

@ -83,6 +83,7 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32, GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW, 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_F32_L4,
GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4_LARGE,
GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32, GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32, GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32, GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32,
@ -533,6 +534,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32, mul_mv_f16_f32, ctx->support_simdgroup_reduction); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32, mul_mv_f16_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW, mul_mv_f16_f32_1row, ctx->support_simdgroup_reduction); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW, mul_mv_f16_f32_1row, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4, mul_mv_f16_f32_l4, ctx->support_simdgroup_reduction); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4, mul_mv_f16_f32_l4, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4_LARGE, mul_mv_f16_f32_l4_large, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32, mul_mv_q4_0_f32, ctx->support_simdgroup_reduction); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32, mul_mv_q4_0_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32, mul_mv_q4_1_f32, ctx->support_simdgroup_reduction); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32, mul_mv_q4_1_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32, mul_mv_q5_0_f32, ctx->support_simdgroup_reduction); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32, mul_mv_q5_0_f32, ctx->support_simdgroup_reduction);
@ -1586,7 +1588,11 @@ static enum ggml_status ggml_metal_graph_compute(
nth1 = 1; nth1 = 1;
if (src1t == GGML_TYPE_F32) { if (src1t == GGML_TYPE_F32) {
if (ne11 * ne12 < 4) { if (ne11 * ne12 < 4) {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW].pipeline; if (ne01 > 128) {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4_LARGE].pipeline;
} else {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4].pipeline;
}
} else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) { } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4].pipeline; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4].pipeline;
nrows = ne11; nrows = ne11;
@ -1778,7 +1784,11 @@ static enum ggml_status ggml_metal_graph_compute(
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
} else { } else {
const int64_t ny = (ne11 + nrows - 1)/nrows; const int64_t ny = (ne11 + nrows - 1)/nrows;
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; if (ne01 > 128) {
[encoder dispatchThreadgroups:MTLSizeMake(ne01/32, ny, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
} else {
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
} }
} }
} break; } break;