root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-26 03:14:35 +00:00
Code Issues Actions 7 Packages Projects Releases Wiki Activity

vulkan: only use M-sized matmul on Apple GPUs (#5412)

Browse Source 
* vulkan: refactor guess_matmul_pipeline for vendor

Refactor ggml_vk_guess_matmul_pipeline to simplify adding per-vendor
conditionals.

Signed-off-by: Sergio Lopez <slp@redhat.com>

* vulkan: only use M-sized matmul on Apple GPUs

L-sized and S-sized matmuls are broken on Apple GPUs, force using
M-size with this vendor.

Signed-off-by: Sergio Lopez <slp@redhat.com>

---------

Signed-off-by: Sergio Lopez <slp@redhat.com>
This commit is contained in:
Sergio López 2024-02-11 15:12:00 +01:00 committed by GitHub
parent a803333a4e
commit c88c74f967
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 93 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

103
ggml-vulkan.cpp
View File

@ -27,6 +27,7 @@
#define CEIL_DIV(M, N) (((M) + (N)-1) / (N)) #define CEIL_DIV(M, N) (((M) + (N)-1) / (N))
#define VK_VENDOR_ID_AMD 0x1002 #define VK_VENDOR_ID_AMD 0x1002
#define VK_VENDOR_ID_APPLE 0x106b
#define VK_VENDOR_ID_INTEL 0x8086 #define VK_VENDOR_ID_INTEL 0x8086
#define VK_VENDOR_ID_NVIDIA 0x10de #define VK_VENDOR_ID_NVIDIA 0x10de
@ -2034,18 +2035,100 @@ static uint32_t ggml_vk_guess_matmul_pipeline_align(ggml_backend_vk_context * ct
return ctx->pipeline_matmul_f32_aligned_l.align; return ctx->pipeline_matmul_f32_aligned_l.align;
} }
static vk_pipeline* ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx, bool bit16_x, bool bit16_y, int m, int n, bool aligned) { static vk_pipeline* ggml_vk_guess_matmul_pipeline_amd(ggml_backend_vk_context * ctx, bool bit16_x, bool bit16_y, int m, int n, bool aligned) {
#ifdef GGML_VULKAN_DEBUG
std::cerr << "ggml_vk_guess_matmul_pipeline(" << bit16_x << ", " << bit16_y << ", " << m << ", " << n << ", " << aligned << ")";
#endif
if (bit16_x && bit16_y) { if (bit16_x && bit16_y) {
if (ctx->device.lock()->vendor_id == VK_VENDOR_ID_INTEL || m <= 32 || n <= 32) { if (m <= 32 || n <= 32) {
#ifdef GGML_VULKAN_DEBUG #ifdef GGML_VULKAN_DEBUG
std::cerr << " S" << std::endl; std::cerr << " S" << std::endl;
#endif #endif
return aligned ? &ctx->pipeline_matmul_f16_aligned_s : &ctx->pipeline_matmul_f16_s; return aligned ? &ctx->pipeline_matmul_f16_aligned_s : &ctx->pipeline_matmul_f16_s;
} }
if (ctx->device.lock()->subgroup_size == 64 || m <= 64 || n <= 64) { #ifdef GGML_VULKAN_DEBUG
std::cerr << " M" << std::endl;
#endif
return aligned ? &ctx->pipeline_matmul_f16_aligned_m : &ctx->pipeline_matmul_f16_m;
}
if (bit16_x && !bit16_y) {
if (m <= 32 || n <= 32) {
#ifdef GGML_VULKAN_DEBUG
std::cerr << " S" << std::endl;
#endif
return aligned ? &ctx->pipeline_matmul_f16_f32_aligned_s : &ctx->pipeline_matmul_f16_f32_s;
}
#ifdef GGML_VULKAN_DEBUG
std::cerr << " M" << std::endl;
#endif
return aligned ? &ctx->pipeline_matmul_f16_f32_aligned_m : &ctx->pipeline_matmul_f16_f32_m;
}
if (!bit16_x && bit16_y) {
GGML_ASSERT(false);
}
if (m <= 32 || n <= 32) {
#ifdef GGML_VULKAN_DEBUG
std::cerr << " S" << std::endl;
#endif
return aligned ? &ctx->pipeline_matmul_f32_aligned_s : &ctx->pipeline_matmul_f32_s;
}
#ifdef GGML_VULKAN_DEBUG
std::cerr << " M" << std::endl;
#endif
return aligned ? &ctx->pipeline_matmul_f32_aligned_m : &ctx->pipeline_matmul_f32_m;
}
static vk_pipeline* ggml_vk_guess_matmul_pipeline_apple(ggml_backend_vk_context * ctx, bool bit16_x, bool bit16_y, bool aligned) {
#ifdef GGML_VULKAN_DEBUG
std::cerr << " M" << std::endl;
#endif
if (bit16_x && bit16_y) {
return aligned ? &ctx->pipeline_matmul_f16_aligned_m : &ctx->pipeline_matmul_f16_m;
}
if (bit16_x && !bit16_y) {
return aligned ? &ctx->pipeline_matmul_f16_f32_aligned_m : &ctx->pipeline_matmul_f16_f32_m;
}
if (!bit16_x && bit16_y) {
GGML_ASSERT(false);
}
return aligned ? &ctx->pipeline_matmul_f32_aligned_m : &ctx->pipeline_matmul_f32_m;
}
static vk_pipeline* ggml_vk_guess_matmul_pipeline_intel(ggml_backend_vk_context * ctx, bool bit16_x, bool bit16_y, bool aligned) {
#ifdef GGML_VULKAN_DEBUG
std::cerr << " S" << std::endl;
#endif
if (bit16_x && bit16_y) {
return aligned ? &ctx->pipeline_matmul_f16_aligned_s : &ctx->pipeline_matmul_f16_s;
}
if (bit16_x && !bit16_y) {
return aligned ? &ctx->pipeline_matmul_f16_f32_aligned_s : &ctx->pipeline_matmul_f16_f32_s;
}
if (!bit16_x && bit16_y) {
GGML_ASSERT(false);
}
return aligned ? &ctx->pipeline_matmul_f32_aligned_s : &ctx->pipeline_matmul_f32_s;
}
static vk_pipeline* ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx, bool bit16_x, bool bit16_y, int m, int n, bool aligned) {
#ifdef GGML_VULKAN_DEBUG
std::cerr << "ggml_vk_guess_matmul_pipeline(" << bit16_x << ", " << bit16_y << ", " << m << ", " << n << ", " << aligned << ")";
#endif
switch (ctx->device.lock()->vendor_id) {
case VK_VENDOR_ID_AMD:
return ggml_vk_guess_matmul_pipeline_amd(ctx, bit16_x, bit16_y, m, n, aligned);
case VK_VENDOR_ID_APPLE:
return ggml_vk_guess_matmul_pipeline_apple(ctx, bit16_x, bit16_y, aligned);
case VK_VENDOR_ID_INTEL:
return ggml_vk_guess_matmul_pipeline_intel(ctx, bit16_x, bit16_y, aligned);
}
if (bit16_x && bit16_y) {
if (m <= 32 || n <= 32) {
#ifdef GGML_VULKAN_DEBUG
std::cerr << " S" << std::endl;
#endif
return aligned ? &ctx->pipeline_matmul_f16_aligned_s : &ctx->pipeline_matmul_f16_s;
}
if (m <= 64 || n <= 64) {
#ifdef GGML_VULKAN_DEBUG #ifdef GGML_VULKAN_DEBUG
std::cerr << " M" << std::endl; std::cerr << " M" << std::endl;
#endif #endif
@ -2057,13 +2140,13 @@ static vk_pipeline* ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx,
return aligned ? &ctx->pipeline_matmul_f16_aligned_l : &ctx->pipeline_matmul_f16_l; return aligned ? &ctx->pipeline_matmul_f16_aligned_l : &ctx->pipeline_matmul_f16_l;
} }
if (bit16_x && !bit16_y) { if (bit16_x && !bit16_y) {
if (ctx->device.lock()->vendor_id == VK_VENDOR_ID_INTEL || m <= 32 || n <= 32) { if (m <= 32 || n <= 32) {
#ifdef GGML_VULKAN_DEBUG #ifdef GGML_VULKAN_DEBUG
std::cerr << " S" << std::endl; std::cerr << " S" << std::endl;
#endif #endif
return aligned ? &ctx->pipeline_matmul_f16_f32_aligned_s : &ctx->pipeline_matmul_f16_f32_s; return aligned ? &ctx->pipeline_matmul_f16_f32_aligned_s : &ctx->pipeline_matmul_f16_f32_s;
} }
if (ctx->device.lock()->subgroup_size == 64 || m <= 64 || n <= 64) { if (m <= 64 || n <= 64) {
#ifdef GGML_VULKAN_DEBUG #ifdef GGML_VULKAN_DEBUG
std::cerr << " M" << std::endl; std::cerr << " M" << std::endl;
#endif #endif
@ -2078,13 +2161,13 @@ static vk_pipeline* ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx,
GGML_ASSERT(false); GGML_ASSERT(false);
} }
if (ctx->device.lock()->vendor_id == VK_VENDOR_ID_INTEL || m <= 32 || n <= 32) { if (m <= 32 || n <= 32) {
#ifdef GGML_VULKAN_DEBUG #ifdef GGML_VULKAN_DEBUG
std::cerr << " S" << std::endl; std::cerr << " S" << std::endl;
#endif #endif
return aligned ? &ctx->pipeline_matmul_f32_aligned_s : &ctx->pipeline_matmul_f32_s; return aligned ? &ctx->pipeline_matmul_f32_aligned_s : &ctx->pipeline_matmul_f32_s;
} }
if (ctx->device.lock()->subgroup_size == 64 || m <= 64 || n <= 64) { if (m <= 64 || n <= 64) {
#ifdef GGML_VULKAN_DEBUG #ifdef GGML_VULKAN_DEBUG
std::cerr << " M" << std::endl; std::cerr << " M" << std::endl;
#endif #endif