vulkan: add support for expect/assume

This commit adds support for compiler hints.
While on AMD these are not used/needed, Nvidia benefits from them, and gives
a sizeable 10% speedup on 4k.
This commit is contained in:
Lynne
2025-04-11 23:44:26 +00:00
parent 985a26be28
commit e040c087c7
6 changed files with 27 additions and 12 deletions

View File

@ -79,6 +79,7 @@ typedef struct VulkanDeviceFeatures {
VkPhysicalDeviceVulkan12Features vulkan_1_2;
VkPhysicalDeviceVulkan13Features vulkan_1_3;
VkPhysicalDeviceTimelineSemaphoreFeatures timeline_semaphore;
VkPhysicalDeviceShaderExpectAssumeFeatures expect_assume;
VkPhysicalDeviceVideoMaintenance1FeaturesKHR video_maintenance_1;
#ifdef VK_KHR_video_maintenance2
@ -210,6 +211,9 @@ static void device_features_init(AVHWDeviceContext *ctx, VulkanDeviceFeatures *f
OPT_CHAIN(&feats->timeline_semaphore, FF_VK_EXT_PORTABILITY_SUBSET,
VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_FEATURES);
OPT_CHAIN(&feats->expect_assume, FF_VK_EXT_EXPECT_ASSUME,
VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_EXPECT_ASSUME_FEATURES_KHR);
OPT_CHAIN(&feats->video_maintenance_1, FF_VK_EXT_VIDEO_MAINTENANCE_1,
VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VIDEO_MAINTENANCE_1_FEATURES_KHR);
#ifdef VK_KHR_video_maintenance2
@ -302,6 +306,8 @@ static void device_features_copy_needed(VulkanDeviceFeatures *dst, VulkanDeviceF
COPY_VAL(relaxed_extended_instruction.shaderRelaxedExtendedInstruction);
#endif
COPY_VAL(expect_assume.shaderExpectAssume);
COPY_VAL(optical_flow.opticalFlow);
#undef COPY_VAL
}
@ -616,6 +622,7 @@ static const VulkanOptExtension optional_device_exts[] = {
{ VK_KHR_COOPERATIVE_MATRIX_EXTENSION_NAME, FF_VK_EXT_COOP_MATRIX },
{ VK_NV_OPTICAL_FLOW_EXTENSION_NAME, FF_VK_EXT_OPTICAL_FLOW },
{ VK_EXT_SHADER_OBJECT_EXTENSION_NAME, FF_VK_EXT_SHADER_OBJECT },
{ VK_KHR_SHADER_EXPECT_ASSUME_EXTENSION_NAME, FF_VK_EXT_EXPECT_ASSUME },
{ VK_KHR_VIDEO_MAINTENANCE_1_EXTENSION_NAME, FF_VK_EXT_VIDEO_MAINTENANCE_1 },
#ifdef VK_KHR_video_maintenance2
{ VK_KHR_VIDEO_MAINTENANCE_2_EXTENSION_NAME, FF_VK_EXT_VIDEO_MAINTENANCE_2 },

View File

@ -2046,6 +2046,12 @@ int ff_vk_shader_init(FFVulkanContext *s, FFVulkanShader *shd, const char *name,
GLSLC(0, #extension GL_EXT_scalar_block_layout : require );
GLSLC(0, #extension GL_EXT_shader_explicit_arithmetic_types : require );
GLSLC(0, #extension GL_EXT_control_flow_attributes : require );
if (s->extensions & FF_VK_EXT_EXPECT_ASSUME) {
GLSLC(0, #extension GL_EXT_expect_assume : require );
} else {
GLSLC(0, #define assumeEXT(x) (x) );
GLSLC(0, #define expectEXT(x, c) (x) );
}
if ((s->extensions & FF_VK_EXT_DEBUG_UTILS) &&
(s->extensions & FF_VK_EXT_RELAXED_EXTENDED_INSTR)) {
GLSLC(0, #extension GL_EXT_debug_printf : require );

View File

@ -47,6 +47,7 @@ typedef uint64_t FFVulkanExtensions;
#define FF_VK_EXT_SHADER_OBJECT (1ULL << 13) /* VK_EXT_shader_object */
#define FF_VK_EXT_PUSH_DESCRIPTOR (1ULL << 14) /* VK_KHR_push_descriptor */
#define FF_VK_EXT_RELAXED_EXTENDED_INSTR (1ULL << 15) /* VK_KHR_shader_relaxed_extended_instruction */
#define FF_VK_EXT_EXPECT_ASSUME (1ULL << 16) /* VK_KHR_shader_expect_assume */
/* Video extensions */
#define FF_VK_EXT_VIDEO_QUEUE (1ULL << 36) /* VK_KHR_video_queue */

View File

@ -76,6 +76,7 @@ static inline uint64_t ff_vk_extensions_to_mask(const char * const *extensions,
{ VK_KHR_VIDEO_DECODE_H265_EXTENSION_NAME, FF_VK_EXT_VIDEO_DECODE_H265 },
{ VK_KHR_VIDEO_DECODE_AV1_EXTENSION_NAME, FF_VK_EXT_VIDEO_DECODE_AV1 },
{ VK_KHR_PUSH_DESCRIPTOR_EXTENSION_NAME, FF_VK_EXT_PUSH_DESCRIPTOR },
{ VK_KHR_SHADER_EXPECT_ASSUME_EXTENSION_NAME, FF_VK_EXT_EXPECT_ASSUME },
};
FFVulkanExtensions mask = 0x0;