mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 32 Packages Projects Releases Wiki Activity

vulkan: Support UPSCALE w/antialias (#18327)

Browse Source
This commit is contained in:
Jeff Bolz
2025-12-26 10:00:57 -06:00
committed by GitHub
parent 10dc500bdb
commit b96b82fc85
3 changed files with 84 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/ggml-vulkan.cpp
+11 -3
View File
@@ -651,7 +651,7 @@ struct vk_device_struct {
vk_pipeline pipeline_add_id_f32;
vk_pipeline pipeline_concat_f32, pipeline_concat_f16, pipeline_concat_i32;
vk_pipeline pipeline_upscale_nearest_f32, pipeline_upscale_bilinear_f32, pipeline_upscale_bicubic_f32;
vk_pipeline pipeline_upscale_nearest_f32, pipeline_upscale_bilinear_f32, pipeline_upscale_bicubic_f32, pipeline_upscale_bilinear_antialias_f32;
vk_pipeline pipeline_scale_f32;
vk_pipeline pipeline_sqr_f32;
vk_pipeline pipeline_sqrt_f32;
@@ -3956,6 +3956,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
ggml_vk_create_pipeline(device, device->pipeline_upscale_nearest_f32, "upscale_f32", upscale_f32_len, upscale_f32_data, "main", 2, sizeof(vk_op_upscale_push_constants), {512, 1, 1}, {GGML_SCALE_MODE_NEAREST}, 1);
ggml_vk_create_pipeline(device, device->pipeline_upscale_bilinear_f32, "upscale_f32", upscale_f32_len, upscale_f32_data, "main", 2, sizeof(vk_op_upscale_push_constants), {512, 1, 1}, {GGML_SCALE_MODE_BILINEAR}, 1);
ggml_vk_create_pipeline(device, device->pipeline_upscale_bicubic_f32, "upscale_f32", upscale_f32_len, upscale_f32_data, "main", 2, sizeof(vk_op_upscale_push_constants), {512, 1, 1}, {GGML_SCALE_MODE_BICUBIC}, 1);
ggml_vk_create_pipeline(device, device->pipeline_upscale_bilinear_antialias_f32, "upscale_f32", upscale_f32_len, upscale_f32_data, "main", 2, sizeof(vk_op_upscale_push_constants), {512, 1, 1}, {GGML_SCALE_MODE_BILINEAR | GGML_SCALE_FLAG_ANTIALIAS}, 1);
ggml_vk_create_pipeline(device, device->pipeline_scale_f32, "scale_f32", scale_f32_len, scale_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
@@ -8433,7 +8434,7 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
return nullptr;
case GGML_OP_UPSCALE:
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
ggml_scale_mode mode = (ggml_scale_mode)(ggml_get_op_params_i32(dst, 0) & 0xFF);
uint32_t mode = (ggml_get_op_params_i32(dst, 0) & (0xFF | GGML_SCALE_FLAG_ANTIALIAS));
switch (mode) {
case GGML_SCALE_MODE_NEAREST:
return ctx->device->pipeline_upscale_nearest_f32;
@@ -8441,6 +8442,8 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
return ctx->device->pipeline_upscale_bilinear_f32;
case GGML_SCALE_MODE_BICUBIC:
return ctx->device->pipeline_upscale_bicubic_f32;
case GGML_SCALE_MODE_BILINEAR | GGML_SCALE_FLAG_ANTIALIAS:
return ctx->device->pipeline_upscale_bilinear_antialias_f32;
default:
return nullptr;
}
@@ -14341,7 +14344,12 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
}
return true;
case GGML_OP_UPSCALE:
return op->src[0]->type == GGML_TYPE_F32 && !(op->op_params[0] & GGML_SCALE_FLAG_ANTIALIAS);
if (op->op_params[0] & GGML_SCALE_FLAG_ANTIALIAS) {
if ((op->op_params[0] & 0xFF) != GGML_SCALE_MODE_BILINEAR) {
return false;
}
}
return op->src[0]->type == GGML_TYPE_F32;
case GGML_OP_ACC:
return op->src[0]->type == GGML_TYPE_F32;
case GGML_OP_CONCAT: