CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-31 08:51:55 +00:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: Skip syncing for prealloc_y when it is reused (#15544)

Browse Source
This commit is contained in:
Jeff Bolz
2025-08-30 04:11:22 -05:00
committed by GitHub
parent ef476916bb
commit 696fccf354
1 changed files with 15 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files

35
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View File

@@ -5800,11 +5800,6 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
ggml_vk_sync_buffers(ctx, subctx);
}
}
if (y_non_contig || quantize_y) {
if (ctx->prealloc_y_need_sync) {
ggml_vk_sync_buffers(ctx, subctx);
}
}
if (x_non_contig) {
ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, { d_Qx, qx_buf_offset, VK_WHOLE_SIZE }, { d_X, 0, VK_WHOLE_SIZE });
@@ -5816,6 +5811,9 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
if (y_non_contig) {
if (ctx->prealloc_y_last_pipeline_used != to_fp16_vk_1.get() ||
ctx->prealloc_y_last_tensor_used != src1) {
if (ctx->prealloc_y_need_sync) {
ggml_vk_sync_buffers(ctx, subctx);
}
ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE });
ctx->prealloc_y_last_pipeline_used = to_fp16_vk_1.get();
ctx->prealloc_y_last_tensor_used = src1;
@@ -5824,6 +5822,9 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
if (quantize_y) {
if (ctx->prealloc_y_last_pipeline_used != to_q8_1.get() ||
ctx->prealloc_y_last_tensor_used != src1) {
if (ctx->prealloc_y_need_sync) {
ggml_vk_sync_buffers(ctx, subctx);
}
ggml_vk_quantize_q8_1(ctx, subctx, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE }, y_ne * ne12 * ne13);
ctx->prealloc_y_last_pipeline_used = to_q8_1.get();
ctx->prealloc_y_last_tensor_used = src1;
@@ -6008,11 +6009,6 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
ggml_vk_sync_buffers(ctx, subctx);
}
}
if (y_non_contig) {
if (ctx->prealloc_y_need_sync) {
ggml_vk_sync_buffers(ctx, subctx);
}
}
if (x_non_contig) {
GGML_ASSERT(x_sz == ggml_vk_align_size(ggml_type_size(src0->type) * x_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment));
@@ -6022,6 +6018,9 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
GGML_ASSERT(y_sz == ggml_type_size(src1->type) * y_ne);
if (ctx->prealloc_y_last_pipeline_used != to_fp16_vk_1.get() ||
ctx->prealloc_y_last_tensor_used != src1) {
if (ctx->prealloc_y_need_sync) {
ggml_vk_sync_buffers(ctx, subctx);
}
ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE });
ctx->prealloc_y_last_pipeline_used = to_fp16_vk_1.get();
ctx->prealloc_y_last_tensor_used = src1;
@@ -6454,11 +6453,6 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
ggml_vk_sync_buffers(ctx, subctx);
}
}
if (y_non_contig) {
if (ctx->prealloc_y_need_sync) {
ggml_vk_sync_buffers(ctx, subctx);
}
}
if (x_non_contig) {
ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, { d_Qx, qx_buf_offset, VK_WHOLE_SIZE }, { d_X, 0, VK_WHOLE_SIZE });
@@ -6471,6 +6465,9 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
if (y_non_contig) {
if (ctx->prealloc_y_last_pipeline_used != to_fp16_vk_1.get() ||
ctx->prealloc_y_last_tensor_used != src1) {
if (ctx->prealloc_y_need_sync) {
ggml_vk_sync_buffers(ctx, subctx);
}
ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE });
ctx->prealloc_y_last_pipeline_used = to_fp16_vk_1.get();
ctx->prealloc_y_last_tensor_used = src1;
@@ -6668,11 +6665,6 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
ggml_vk_sync_buffers(ctx, subctx);
}
}
if (y_non_contig) {
if (ctx->prealloc_y_need_sync) {
ggml_vk_sync_buffers(ctx, subctx);
}
}
if (x_non_contig) {
GGML_ASSERT(x_sz == ggml_vk_align_size(ggml_type_size(src0->type) * x_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment));
@@ -6682,6 +6674,9 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
GGML_ASSERT(y_sz == ggml_type_size(src1->type) * y_ne);
if (ctx->prealloc_y_last_pipeline_used != to_fp16_vk_1.get() ||
ctx->prealloc_y_last_tensor_used != src1) {
if (ctx->prealloc_y_need_sync) {
ggml_vk_sync_buffers(ctx, subctx);
}
ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE });
ctx->prealloc_y_last_pipeline_used = to_fp16_vk_1.get();
ctx->prealloc_y_last_tensor_used = src1;