tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-27 20:05:20 +00:00
Code Issues Packages Projects Releases Wiki Activity

sync : ggml (#13268)

Browse Source 
* vulkan : kernels for depthwise 2D convolution (CONV_2D_DW) (ggml/1204)

* vulkan : add kernels for depthwise 2d convolution (OP_CONV_2D_DW)

* review: remove src_x/y < 0 checks; add performance tests

* sync : ggml

ggml-ci

* vulkan : fix lint (#0)

---------

Co-authored-by: Acly <aclysia@gmail.com>
This commit is contained in:
Georgi Gerganov
2025-05-02 20:54:30 +03:00
committed by GitHub
parent a75cb30dc9
commit b34443923c
5 changed files with 225 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

66
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View File

@ -389,6 +389,8 @@ struct vk_device_struct {
vk_pipeline pipeline_rwkv_wkv6_f32;
vk_pipeline pipeline_rwkv_wkv7_f32;
vk_pipeline pipeline_opt_step_adamw_f32;
vk_pipeline pipeline_conv2d_dw_whcn_f32;
vk_pipeline pipeline_conv2d_dw_cwhn_f32;
// [2][2][2] is for {f16acc,f32acc}x{large,small_rows}x{unaligned, aligned}
vk_pipeline pipeline_flash_attn_f32_f16_D64[GGML_TYPE_COUNT][2][2][2];
@ -701,6 +703,24 @@ struct vk_op_rwkv_wkv7_push_constants {
uint32_t H;
};
struct vk_op_conv2d_dw_push_constants {
uint32_t ne;
uint32_t batches;
uint32_t channels;
uint32_t dst_w;
uint32_t dst_h;
uint32_t src_w;
uint32_t src_h;
uint32_t knl_w;
uint32_t knl_h;
int32_t stride_x;
int32_t stride_y;
int32_t pad_x;
int32_t pad_y;
int32_t dilation_x;
int32_t dilation_y;
};
struct vk_op_upscale_push_constants {
uint32_t ne; uint32_t a_offset; uint32_t d_offset;
uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03;
@ -2610,6 +2630,9 @@ static void ggml_vk_load_shaders(vk_device& device) {
ggml_vk_create_pipeline(device, device->pipeline_opt_step_adamw_f32, "opt_step_adamw_f32", opt_step_adamw_f32_len, opt_step_adamw_f32_data, "main", 5, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_conv2d_dw_whcn_f32, "conv2d_dw_whcn_f32", conv2d_dw_whcn_f32_len, conv2d_dw_whcn_f32_data, "main", 3, sizeof(vk_op_conv2d_dw_push_constants), {512, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_conv2d_dw_cwhn_f32, "conv2d_dw_cwhn_f32", conv2d_dw_cwhn_f32_len, conv2d_dw_cwhn_f32_data, "main", 3, sizeof(vk_op_conv2d_dw_push_constants), {512, 1, 1}, {}, 1);
for (auto &c : compiles) {
c.wait();
}
@ -6137,6 +6160,15 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
return ctx->device->pipeline_leaky_relu_f32;
}
return nullptr;
case GGML_OP_CONV_2D_DW:
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
if (ggml_is_contiguous(src1)) {
return ctx->device->pipeline_conv2d_dw_whcn_f32;
} else if (ggml_is_contiguous_channels(src1)) {
return ctx->device->pipeline_conv2d_dw_cwhn_f32;
}
}
return nullptr;
default:
return nullptr;
}
@ -6163,6 +6195,7 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
case GGML_OP_REPEAT_BACK:
case GGML_OP_ROPE:
case GGML_OP_RMS_NORM:
case GGML_OP_CONV_2D_DW:
return true;
default:
return false;
@ -6459,6 +6492,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
case GGML_OP_CONCAT:
case GGML_OP_UPSCALE:
case GGML_OP_UNARY:
case GGML_OP_CONV_2D_DW:
{
const uint32_t ne = ggml_nelements(dst);
if (ne > 262144) {
@ -7245,6 +7279,30 @@ static void ggml_vk_pool_2d(ggml_backend_vk_context * ctx, vk_context& subctx, c
}, dryrun);
}
static void ggml_vk_conv_2d_dw(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
vk_op_conv2d_dw_push_constants p{};
p.ne = ggml_nelements(dst);
p.channels = dst->ne[2];
p.batches = dst->ne[3];
p.dst_w = dst->ne[0];
p.dst_h = dst->ne[1];
p.src_w = src1->ne[0];
p.src_h = src1->ne[1];
p.knl_w = src0->ne[0];
p.knl_h = src0->ne[1];
p.stride_x = dst->op_params[0];
p.stride_y = dst->op_params[1];
p.pad_x = dst->op_params[2];
p.pad_y = dst->op_params[3];
p.dilation_x = dst->op_params[4];
p.dilation_y = dst->op_params[5];
GGML_ASSERT(src0->ne[3] == p.channels);
GGML_ASSERT(src1->ne[3] == p.batches);
ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_CONV_2D_DW, std::move(p), dryrun);
}
static void ggml_vk_leaky_relu(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
const float * op_params = (const float *)dst->op_params;
ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_LEAKY_RELU, { (uint32_t)ggml_nelements(src0), 0, op_params[0], 0.0f }, dryrun);
@ -8265,6 +8323,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
case GGML_OP_IM2COL:
case GGML_OP_TIMESTEP_EMBEDDING:
case GGML_OP_POOL_2D:
case GGML_OP_CONV_2D_DW:
case GGML_OP_RWKV_WKV6:
case GGML_OP_RWKV_WKV7:
case GGML_OP_LEAKY_RELU:
@ -8328,6 +8387,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
case GGML_OP_IM2COL:
case GGML_OP_TIMESTEP_EMBEDDING:
case GGML_OP_POOL_2D:
case GGML_OP_CONV_2D_DW:
case GGML_OP_LEAKY_RELU:
{
// These operations all go through ggml_vk_op_f32, so short-circuit and
@ -8501,6 +8561,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
case GGML_OP_POOL_2D:
ggml_vk_pool_2d(ctx, compute_ctx, src0, node, dryrun);
break;
case GGML_OP_CONV_2D_DW:
ggml_vk_conv_2d_dw(ctx, compute_ctx, src0, src1, node, dryrun);
break;
case GGML_OP_LEAKY_RELU:
ggml_vk_leaky_relu(ctx, compute_ctx, src0, node, dryrun);
@ -8622,6 +8686,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
case GGML_OP_IM2COL:
case GGML_OP_TIMESTEP_EMBEDDING:
case GGML_OP_POOL_2D:
case GGML_OP_CONV_2D_DW:
case GGML_OP_RWKV_WKV6:
case GGML_OP_RWKV_WKV7:
case GGML_OP_LEAKY_RELU:
@ -9599,6 +9664,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
case GGML_OP_COUNT_EQUAL:
case GGML_OP_IM2COL:
case GGML_OP_TIMESTEP_EMBEDDING:
case GGML_OP_CONV_2D_DW:
case GGML_OP_POOL_2D:
case GGML_OP_RWKV_WKV6:
case GGML_OP_RWKV_WKV7: