radv/video: add initial support for encoding with h264.

This adds the encoding infrastructure along with support for h264.

Acked-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25900>
This commit is contained in:
Dave Airlie 2023-10-25 15:56:16 +10:00
parent 800c03ffbd
commit 54d499818c
6 changed files with 1635 additions and 7 deletions

View file

@ -162,6 +162,7 @@ libradv_files = files(
'radv_query.h', 'radv_query.h',
'radv_video.c', 'radv_video.c',
'radv_video.h', 'radv_video.h',
'radv_video_enc.c',
'radv_wsi.c', 'radv_wsi.c',
'radv_wsi.h', 'radv_wsi.h',
) )

View file

@ -437,6 +437,21 @@ struct radv_cmd_state {
bool uses_dynamic_vertex_binding_stride; bool uses_dynamic_vertex_binding_stride;
}; };
struct radv_enc_state {
uint32_t task_size_offset;
uint32_t total_task_size;
unsigned shifter;
unsigned bits_in_shifter;
uint32_t num_zeros;
uint32_t byte_index;
unsigned bits_output;
unsigned bits_size;
bool emulation_prevention;
bool is_even_frame;
unsigned task_id;
uint32_t copy_start_offset;
};
struct radv_cmd_buffer_upload { struct radv_cmd_buffer_upload {
uint8_t *map; uint8_t *map;
unsigned offset; unsigned offset;
@ -537,6 +552,7 @@ struct radv_cmd_buffer {
struct radv_video_session_params *params; struct radv_video_session_params *params;
struct rvcn_sq_var sq; struct rvcn_sq_var sq;
struct rvcn_decode_buffer_s *decode_buffer; struct rvcn_decode_buffer_s *decode_buffer;
struct radv_enc_state enc;
uint64_t feedback_query_va; uint64_t feedback_query_va;
} video; } video;

View file

@ -17,7 +17,7 @@
#include "radv_instance.h" #include "radv_instance.h"
#include "radv_queue.h" #include "radv_queue.h"
#include "radv_radeon_winsys.h" #include "radv_radeon_winsys.h"
#include "ac_vcn_enc.h"
#include "wsi_common.h" #include "wsi_common.h"
#include "nir.h" #include "nir.h"
@ -69,6 +69,13 @@ struct radv_physical_device_cache_key {
uint32_t use_ngg_culling : 1; uint32_t use_ngg_culling : 1;
}; };
enum radv_video_enc_hw_ver {
RADV_VIDEO_ENC_HW_1_2,
RADV_VIDEO_ENC_HW_2,
RADV_VIDEO_ENC_HW_3,
RADV_VIDEO_ENC_HW_4,
};
struct radv_physical_device { struct radv_physical_device {
struct vk_physical_device vk; struct vk_physical_device vk;
@ -171,6 +178,9 @@ struct radv_physical_device {
uint32_t stream_handle_base; uint32_t stream_handle_base;
uint32_t stream_handle_counter; uint32_t stream_handle_counter;
uint32_t av1_version; uint32_t av1_version;
rvcn_enc_cmd_t vcn_enc_cmds;
enum radv_video_enc_hw_ver enc_hw_ver;
uint32_t encoder_interface_version;
struct radv_physical_device_cache_key cache_key; struct radv_physical_device_cache_key cache_key;
}; };

View file

@ -337,6 +337,9 @@ radv_video_patch_session_parameters(struct vk_video_session_parameters *params)
case VK_VIDEO_CODEC_OPERATION_DECODE_H265_BIT_KHR: case VK_VIDEO_CODEC_OPERATION_DECODE_H265_BIT_KHR:
default: default:
return; return;
case VK_VIDEO_CODEC_OPERATION_ENCODE_H264_BIT_KHR:
radv_video_patch_encode_session_parameters(params);
break;
} }
} }
@ -379,6 +382,31 @@ radv_CreateVideoSessionKHR(VkDevice _device, const VkVideoSessionCreateInfoKHR *
vid->stream_type = RDECODE_CODEC_AV1; vid->stream_type = RDECODE_CODEC_AV1;
vid->dpb_type = DPB_DYNAMIC_TIER_2; vid->dpb_type = DPB_DYNAMIC_TIER_2;
break; break;
case VK_VIDEO_CODEC_OPERATION_ENCODE_H264_BIT_KHR:
vid->encode = true;
vid->enc_session.encode_standard = RENCODE_ENCODE_STANDARD_H264;
vid->enc_session.aligned_picture_width = align(vid->vk.max_coded.width, 16);
vid->enc_session.aligned_picture_height = align(vid->vk.max_coded.height, 16);
vid->enc_session.padding_width = vid->enc_session.aligned_picture_width - vid->vk.max_coded.width;
vid->enc_session.padding_height = vid->enc_session.aligned_picture_height - vid->vk.max_coded.height;
vid->enc_session.display_remote = 0;
vid->enc_session.pre_encode_mode = 0;
vid->enc_session.pre_encode_chroma_enabled = 0;
switch (vid->vk.enc_usage.tuning_mode) {
case VK_VIDEO_ENCODE_TUNING_MODE_DEFAULT_KHR:
default:
vid->enc_preset_mode = RENCODE_PRESET_MODE_BALANCE;
break;
case VK_VIDEO_ENCODE_TUNING_MODE_LOW_LATENCY_KHR:
case VK_VIDEO_ENCODE_TUNING_MODE_ULTRA_LOW_LATENCY_KHR:
vid->enc_preset_mode = RENCODE_PRESET_MODE_SPEED;
break;
case VK_VIDEO_ENCODE_TUNING_MODE_HIGH_QUALITY_KHR:
case VK_VIDEO_ENCODE_TUNING_MODE_LOSSLESS_KHR:
vid->enc_preset_mode = RENCODE_PRESET_MODE_QUALITY;
break;
}
break;
default: default:
return VK_ERROR_FEATURE_NOT_PRESENT; return VK_ERROR_FEATURE_NOT_PRESENT;
} }
@ -451,6 +479,7 @@ radv_GetPhysicalDeviceVideoCapabilitiesKHR(VkPhysicalDevice physicalDevice, cons
{ {
VK_FROM_HANDLE(radv_physical_device, pdev, physicalDevice); VK_FROM_HANDLE(radv_physical_device, pdev, physicalDevice);
const struct video_codec_cap *cap = NULL; const struct video_codec_cap *cap = NULL;
bool is_encode = false;
switch (pVideoProfile->videoCodecOperation) { switch (pVideoProfile->videoCodecOperation) {
#ifndef _WIN32 #ifndef _WIN32
@ -463,6 +492,10 @@ radv_GetPhysicalDeviceVideoCapabilitiesKHR(VkPhysicalDevice physicalDevice, cons
case VK_VIDEO_CODEC_OPERATION_DECODE_AV1_BIT_KHR: case VK_VIDEO_CODEC_OPERATION_DECODE_AV1_BIT_KHR:
cap = &pdev->info.dec_caps.codec_info[AMDGPU_INFO_VIDEO_CAPS_CODEC_IDX_AV1]; cap = &pdev->info.dec_caps.codec_info[AMDGPU_INFO_VIDEO_CAPS_CODEC_IDX_AV1];
break; break;
case VK_VIDEO_CODEC_OPERATION_ENCODE_H264_BIT_KHR:
cap = &pdev->info.enc_caps.codec_info[AMDGPU_INFO_VIDEO_CAPS_CODEC_IDX_MPEG4_AVC];
is_encode = true;
break;
#endif #endif
default: default:
unreachable("unsupported operation"); unreachable("unsupported operation");
@ -472,17 +505,40 @@ radv_GetPhysicalDeviceVideoCapabilitiesKHR(VkPhysicalDevice physicalDevice, cons
cap = NULL; cap = NULL;
pCapabilities->flags = 0; pCapabilities->flags = 0;
pCapabilities->minBitstreamBufferOffsetAlignment = 128;
pCapabilities->minBitstreamBufferSizeAlignment = 128;
pCapabilities->pictureAccessGranularity.width = VL_MACROBLOCK_WIDTH; pCapabilities->pictureAccessGranularity.width = VL_MACROBLOCK_WIDTH;
pCapabilities->pictureAccessGranularity.height = VL_MACROBLOCK_HEIGHT; pCapabilities->pictureAccessGranularity.height = VL_MACROBLOCK_HEIGHT;
pCapabilities->minCodedExtent.width = VL_MACROBLOCK_WIDTH; pCapabilities->minCodedExtent.width = VL_MACROBLOCK_WIDTH;
pCapabilities->minCodedExtent.height = VL_MACROBLOCK_HEIGHT; pCapabilities->minCodedExtent.height = VL_MACROBLOCK_HEIGHT;
struct VkVideoDecodeCapabilitiesKHR *dec_caps = struct VkVideoDecodeCapabilitiesKHR *dec_caps = NULL;
(struct VkVideoDecodeCapabilitiesKHR *)vk_find_struct(pCapabilities->pNext, VIDEO_DECODE_CAPABILITIES_KHR); struct VkVideoEncodeCapabilitiesKHR *enc_caps = NULL;
if (dec_caps) if (!is_encode) {
dec_caps->flags = VK_VIDEO_DECODE_CAPABILITY_DPB_AND_OUTPUT_DISTINCT_BIT_KHR; dec_caps =
(struct VkVideoDecodeCapabilitiesKHR *)vk_find_struct(pCapabilities->pNext, VIDEO_DECODE_CAPABILITIES_KHR);
if (dec_caps)
dec_caps->flags = VK_VIDEO_DECODE_CAPABILITY_DPB_AND_OUTPUT_DISTINCT_BIT_KHR;
pCapabilities->minBitstreamBufferOffsetAlignment = 128;
pCapabilities->minBitstreamBufferSizeAlignment = 128;
} else {
enc_caps =
(struct VkVideoEncodeCapabilitiesKHR *)vk_find_struct(pCapabilities->pNext, VIDEO_ENCODE_CAPABILITIES_KHR);
if (enc_caps) {
enc_caps->flags = 0;
enc_caps->rateControlModes = VK_VIDEO_ENCODE_RATE_CONTROL_MODE_DISABLED_BIT_KHR |
VK_VIDEO_ENCODE_RATE_CONTROL_MODE_CBR_BIT_KHR |
VK_VIDEO_ENCODE_RATE_CONTROL_MODE_VBR_BIT_KHR;
enc_caps->maxRateControlLayers = RADV_ENC_MAX_RATE_LAYER;
enc_caps->maxBitrate = 0;
enc_caps->maxQualityLevels = 2;
enc_caps->encodeInputPictureGranularity.width = 1;
enc_caps->encodeInputPictureGranularity.height = 1;
enc_caps->supportedEncodeFeedbackFlags = VK_VIDEO_ENCODE_FEEDBACK_BITSTREAM_BUFFER_OFFSET_BIT_KHR |
VK_VIDEO_ENCODE_FEEDBACK_BITSTREAM_BYTES_WRITTEN_BIT_KHR;
}
pCapabilities->minBitstreamBufferOffsetAlignment = 16;
pCapabilities->minBitstreamBufferSizeAlignment = 16;
}
switch (pVideoProfile->videoCodecOperation) { switch (pVideoProfile->videoCodecOperation) {
case VK_VIDEO_CODEC_OPERATION_DECODE_H264_BIT_KHR: { case VK_VIDEO_CODEC_OPERATION_DECODE_H264_BIT_KHR: {
@ -573,6 +629,33 @@ radv_GetPhysicalDeviceVideoCapabilitiesKHR(VkPhysicalDevice physicalDevice, cons
pCapabilities->stdHeaderVersion.specVersion = VK_STD_VULKAN_VIDEO_CODEC_AV1_DECODE_SPEC_VERSION; pCapabilities->stdHeaderVersion.specVersion = VK_STD_VULKAN_VIDEO_CODEC_AV1_DECODE_SPEC_VERSION;
break; break;
} }
case VK_VIDEO_CODEC_OPERATION_ENCODE_H264_BIT_KHR: {
struct VkVideoEncodeH264CapabilitiesKHR *ext = (struct VkVideoEncodeH264CapabilitiesKHR *)vk_find_struct(
pCapabilities->pNext, VIDEO_ENCODE_H264_CAPABILITIES_KHR);
pCapabilities->maxDpbSlots = NUM_H2645_REFS;
pCapabilities->maxActiveReferencePictures = NUM_H2645_REFS;
ext->flags = VK_VIDEO_ENCODE_H264_CAPABILITY_HRD_COMPLIANCE_BIT_KHR;
ext->maxLevelIdc = cap ? cap->max_level : 0;
ext->maxSliceCount = 128;
ext->maxPPictureL0ReferenceCount = 1;
ext->maxBPictureL0ReferenceCount = 0;
ext->maxL1ReferenceCount = 0;
ext->maxTemporalLayerCount = 4;
ext->expectDyadicTemporalLayerPattern = false;
ext->minQp = 0;
ext->maxQp = 51;
ext->prefersGopRemainingFrames = false;
ext->requiresGopRemainingFrames = false;
ext->stdSyntaxFlags = VK_VIDEO_ENCODE_H264_STD_CONSTRAINED_INTRA_PRED_FLAG_SET_BIT_KHR |
VK_VIDEO_ENCODE_H264_STD_ENTROPY_CODING_MODE_FLAG_UNSET_BIT_KHR |
VK_VIDEO_ENCODE_H264_STD_ENTROPY_CODING_MODE_FLAG_SET_BIT_KHR;
if (pdev->enc_hw_ver >= RADV_VIDEO_ENC_HW_3)
ext->stdSyntaxFlags |= VK_VIDEO_ENCODE_H264_STD_WEIGHTED_BIPRED_IDC_EXPLICIT_BIT_KHR;
strcpy(pCapabilities->stdHeaderVersion.extensionName, VK_STD_VULKAN_VIDEO_CODEC_H264_ENCODE_EXTENSION_NAME);
pCapabilities->stdHeaderVersion.specVersion = VK_STD_VULKAN_VIDEO_CODEC_H264_ENCODE_SPEC_VERSION;
break;
}
default: default:
break; break;
} }
@ -666,6 +749,10 @@ radv_GetVideoSessionMemoryRequirementsKHR(VkDevice _device, VkVideoSessionKHR vi
uint32_t memory_type_bits = (1u << pdev->memory_properties.memoryTypeCount) - 1; uint32_t memory_type_bits = (1u << pdev->memory_properties.memoryTypeCount) - 1;
if (vid->encode) {
return radv_video_get_encode_session_memory_requirements(device, vid, pMemoryRequirementsCount,
pMemoryRequirements);
}
VK_OUTARRAY_MAKE_TYPED(VkVideoSessionMemoryRequirementsKHR, out, pMemoryRequirements, pMemoryRequirementsCount); VK_OUTARRAY_MAKE_TYPED(VkVideoSessionMemoryRequirementsKHR, out, pMemoryRequirements, pMemoryRequirementsCount);
/* 1 buffer for session context */ /* 1 buffer for session context */
if (pdev->info.family >= CHIP_POLARIS10) { if (pdev->info.family >= CHIP_POLARIS10) {
@ -2751,6 +2838,9 @@ radv_CmdBeginVideoCodingKHR(VkCommandBuffer commandBuffer, const VkVideoBeginCod
cmd_buffer->video.vid = vid; cmd_buffer->video.vid = vid;
cmd_buffer->video.params = params; cmd_buffer->video.params = params;
if (vid->encode)
radv_video_enc_begin_coding(cmd_buffer);
} }
static void static void
@ -2833,6 +2923,10 @@ radv_CmdControlVideoCodingKHR(VkCommandBuffer commandBuffer, const VkVideoCoding
struct radv_device *device = radv_cmd_buffer_device(cmd_buffer); struct radv_device *device = radv_cmd_buffer_device(cmd_buffer);
struct radv_physical_device *pdev = radv_device_physical(device); struct radv_physical_device *pdev = radv_device_physical(device);
if (cmd_buffer->video.vid->encode) {
radv_video_enc_control_video_coding(cmd_buffer, pCodingControlInfo);
return;
}
if (pCodingControlInfo->flags & VK_VIDEO_CODING_CONTROL_RESET_BIT_KHR) { if (pCodingControlInfo->flags & VK_VIDEO_CODING_CONTROL_RESET_BIT_KHR) {
if (radv_has_uvd(pdev)) if (radv_has_uvd(pdev))
radv_uvd_cmd_reset(cmd_buffer); radv_uvd_cmd_reset(cmd_buffer);
@ -2844,6 +2938,12 @@ radv_CmdControlVideoCodingKHR(VkCommandBuffer commandBuffer, const VkVideoCoding
VKAPI_ATTR void VKAPI_CALL VKAPI_ATTR void VKAPI_CALL
radv_CmdEndVideoCodingKHR(VkCommandBuffer commandBuffer, const VkVideoEndCodingInfoKHR *pEndCodingInfo) radv_CmdEndVideoCodingKHR(VkCommandBuffer commandBuffer, const VkVideoEndCodingInfoKHR *pEndCodingInfo)
{ {
VK_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
if (cmd_buffer->video.vid->encode) {
radv_video_enc_end_coding(cmd_buffer);
return;
}
} }
static void static void

View file

@ -13,11 +13,16 @@
#include "vk_video.h" #include "vk_video.h"
#include "ac_vcn.h"
#define VL_MACROBLOCK_WIDTH 16 #define VL_MACROBLOCK_WIDTH 16
#define VL_MACROBLOCK_HEIGHT 16 #define VL_MACROBLOCK_HEIGHT 16
struct radv_physical_device; struct radv_physical_device;
struct rvcn_sq_var; struct rvcn_sq_var;
struct radv_cmd_buffer;
#define RADV_ENC_MAX_RATE_LAYER 4
struct radv_vid_mem { struct radv_vid_mem {
struct radv_device_memory *mem; struct radv_device_memory *mem;
@ -31,6 +36,7 @@ struct radv_video_session {
uint32_t stream_handle; uint32_t stream_handle;
unsigned stream_type; unsigned stream_type;
bool interlaced; bool interlaced;
bool encode;
enum { DPB_MAX_RES = 0, DPB_DYNAMIC_TIER_1, DPB_DYNAMIC_TIER_2 } dpb_type; enum { DPB_MAX_RES = 0, DPB_DYNAMIC_TIER_1, DPB_DYNAMIC_TIER_2 } dpb_type;
unsigned db_alignment; unsigned db_alignment;
@ -38,6 +44,13 @@ struct radv_video_session {
struct radv_vid_mem ctx; struct radv_vid_mem ctx;
unsigned dbg_frame_cnt; unsigned dbg_frame_cnt;
rvcn_enc_session_init_t enc_session;
rvcn_enc_layer_control_t rc_layer_control;
rvcn_enc_rate_ctl_layer_init_t rc_layer_init[RADV_ENC_MAX_RATE_LAYER];
rvcn_enc_rate_ctl_per_picture_t rc_per_pic[RADV_ENC_MAX_RATE_LAYER];
uint32_t enc_preset_mode;
uint32_t enc_rate_control_method;
bool enc_rate_control_default;
}; };
VK_DEFINE_NONDISP_HANDLE_CASTS(radv_video_session, vk.base, VkVideoSessionKHR, VK_OBJECT_TYPE_VIDEO_SESSION_KHR) VK_DEFINE_NONDISP_HANDLE_CASTS(radv_video_session, vk.base, VkVideoSessionKHR, VK_OBJECT_TYPE_VIDEO_SESSION_KHR)
@ -58,4 +71,14 @@ void radv_vcn_sq_header(struct radeon_cmdbuf *cs, struct rvcn_sq_var *sq, bool e
void radv_vcn_sq_tail(struct radeon_cmdbuf *cs, struct rvcn_sq_var *sq); void radv_vcn_sq_tail(struct radeon_cmdbuf *cs, struct rvcn_sq_var *sq);
void radv_init_physical_device_encoder(struct radv_physical_device *pdevice);
void radv_video_enc_begin_coding(struct radv_cmd_buffer *cmd_buffer);
void radv_video_enc_end_coding(struct radv_cmd_buffer *cmd_buffer);
void radv_video_enc_control_video_coding(struct radv_cmd_buffer *cmd_buffer,
const VkVideoCodingControlInfoKHR *pCodingControlInfo);
VkResult radv_video_get_encode_session_memory_requirements(struct radv_device *device, struct radv_video_session *vid,
uint32_t *pMemoryRequirementsCount,
VkVideoSessionMemoryRequirementsKHR *pMemoryRequirements);
void radv_video_patch_encode_session_parameters(struct vk_video_session_parameters *params);
#endif /* RADV_VIDEO_H */ #endif /* RADV_VIDEO_H */

File diff suppressed because it is too large Load diff