gallium/radeon: fix PIPE_QUERY_GPU_FINISHED
[mesa.git] / src / gallium / drivers / radeon / radeon_vce_40_2_2.c
index 3b67b31fdf86a8c6b257d30b0cec02907bbf3cd1..c00565904270a0a063e5736af2e0250b83367afa 100644 (file)
 
 #include "vl/vl_video_buffer.h"
 
-#include "../../winsys/radeon/drm/radeon_winsys.h"
+#include "r600_pipe_common.h"
 #include "radeon_video.h"
 #include "radeon_vce.h"
 
-static struct rvce_cpb_slot *current_slot(struct rvce_encoder *enc)
-{
-       return LIST_ENTRY(struct rvce_cpb_slot, enc->cpb_slots.prev, list);
-}
-
-static struct rvce_cpb_slot *l0_slot(struct rvce_encoder *enc)
-{
-       return LIST_ENTRY(struct rvce_cpb_slot, enc->cpb_slots.next, list);
-}
-
-static struct rvce_cpb_slot *l1_slot(struct rvce_encoder *enc)
-{
-       return LIST_ENTRY(struct rvce_cpb_slot, enc->cpb_slots.next->next, list);
-}
-
-static void frame_offset(struct rvce_encoder *enc, struct rvce_cpb_slot *slot,
-                        unsigned *luma_offset, unsigned *chroma_offset)
-{
-       unsigned pitch = align(enc->luma->level[0].pitch_bytes, 128);
-       unsigned vpitch = align(enc->luma->npix_y, 16);
-       unsigned fsize = pitch * (vpitch + vpitch / 2);
-
-       *luma_offset = slot->index * fsize;
-       *chroma_offset = *luma_offset + pitch * vpitch;
-}
+static const unsigned profiles[7] = { 66, 77, 88, 100, 110, 122, 244 };
 
 static void session(struct rvce_encoder *enc)
 {
@@ -77,35 +53,44 @@ static void session(struct rvce_encoder *enc)
        RVCE_END();
 }
 
-static void task_info(struct rvce_encoder *enc, uint32_t taskOperation)
+static void task_info(struct rvce_encoder *enc, uint32_t op,
+                     uint32_t dep, uint32_t fb_idx, uint32_t ring_idx)
 {
        RVCE_BEGIN(0x00000002); // task info
+       if (op == 0x3) {
+               if (enc->task_info_idx) {
+                       uint32_t offs = enc->cs->cdw - enc->task_info_idx + 3;
+                       // Update offsetOfNextTaskInfo
+                       enc->cs->buf[enc->task_info_idx] = offs;
+               }
+               enc->task_info_idx = enc->cs->cdw;
+       }
        RVCE_CS(0xffffffff); // offsetOfNextTaskInfo
-       RVCE_CS(taskOperation); // taskOperation
-       RVCE_CS(0x00000000); // referencePictureDependency
+       RVCE_CS(op); // taskOperation
+       RVCE_CS(dep); // referencePictureDependency
        RVCE_CS(0x00000000); // collocateFlagDependency
-       RVCE_CS(0x00000000); // feedbackIndex
-       RVCE_CS(0x00000000); // videoBitstreamRingIndex
+       RVCE_CS(fb_idx); // feedbackIndex
+       RVCE_CS(ring_idx); // videoBitstreamRingIndex
        RVCE_END();
 }
 
 static void feedback(struct rvce_encoder *enc)
 {
        RVCE_BEGIN(0x05000005); // feedback buffer
-       RVCE_WRITE(enc->fb->cs_handle, enc->fb->domain); // feedbackRingAddressHi
-       RVCE_CS(0x00000000); // feedbackRingAddressLo
+       RVCE_WRITE(enc->fb->res->cs_buf, enc->fb->res->domains, 0x0); // feedbackRingAddressHi/Lo
        RVCE_CS(0x00000001); // feedbackRingSize
        RVCE_END();
 }
 
 static void create(struct rvce_encoder *enc)
 {
-       task_info(enc, 0x00000000);
+       enc->task_info(enc, 0x00000000, 0, 0, 0);
 
        RVCE_BEGIN(0x01000001); // create cmd
        RVCE_CS(0x00000000); // encUseCircularBuffer
-       RVCE_CS(0x0000004d); // encProfile: Main
-       RVCE_CS(0x0000002a); // encLevel: 4.2
+       RVCE_CS(profiles[enc->base.profile -
+               PIPE_VIDEO_PROFILE_MPEG4_AVC_BASELINE]); // encProfile
+       RVCE_CS(enc->base.level); // encLevel
        RVCE_CS(0x00000000); // encPicStructRestriction
        RVCE_CS(enc->base.width); // encImageWidth
        RVCE_CS(enc->base.height); // encImageHeight
@@ -244,21 +229,85 @@ static void rdo(struct rvce_encoder *enc)
        RVCE_END();
 }
 
+static void vui(struct rvce_encoder *enc)
+{
+       int i;
+
+       if (!enc->pic.rate_ctrl.frame_rate_num)
+               return;
+
+       RVCE_BEGIN(0x04000009); // vui
+       RVCE_CS(0x00000000); //aspectRatioInfoPresentFlag
+       RVCE_CS(0x00000000); //aspectRatioInfo.aspectRatioIdc
+       RVCE_CS(0x00000000); //aspectRatioInfo.sarWidth
+       RVCE_CS(0x00000000); //aspectRatioInfo.sarHeight
+       RVCE_CS(0x00000000); //overscanInfoPresentFlag
+       RVCE_CS(0x00000000); //overScanInfo.overscanAppropFlag
+       RVCE_CS(0x00000000); //videoSignalTypePresentFlag
+       RVCE_CS(0x00000005); //videoSignalTypeInfo.videoFormat
+       RVCE_CS(0x00000000); //videoSignalTypeInfo.videoFullRangeFlag
+       RVCE_CS(0x00000000); //videoSignalTypeInfo.colorDescriptionPresentFlag
+       RVCE_CS(0x00000002); //videoSignalTypeInfo.colorPrim
+       RVCE_CS(0x00000002); //videoSignalTypeInfo.transferChar
+       RVCE_CS(0x00000002); //videoSignalTypeInfo.matrixCoef
+       RVCE_CS(0x00000000); //chromaLocInfoPresentFlag
+       RVCE_CS(0x00000000); //chromaLocInfo.chromaLocTop
+       RVCE_CS(0x00000000); //chromaLocInfo.chromaLocBottom
+       RVCE_CS(0x00000001); //timingInfoPresentFlag
+       RVCE_CS(enc->pic.rate_ctrl.frame_rate_den); //timingInfo.numUnitsInTick
+       RVCE_CS(enc->pic.rate_ctrl.frame_rate_num * 2); //timingInfo.timeScale;
+       RVCE_CS(0x00000001); //timingInfo.fixedFrameRateFlag
+       RVCE_CS(0x00000000); //nalHRDParametersPresentFlag
+       RVCE_CS(0x00000000); //hrdParam.cpbCntMinus1
+       RVCE_CS(0x00000004); //hrdParam.bitRateScale
+       RVCE_CS(0x00000006); //hrdParam.cpbSizeScale
+       for (i = 0; i < 32; i++) {
+               RVCE_CS(0x00000000); //hrdParam.bitRateValueMinus
+               RVCE_CS(0x00000000); //hrdParam.cpbSizeValueMinus
+               RVCE_CS(0x00000000); //hrdParam.cbrFlag
+       }
+       RVCE_CS(0x00000017); //hrdParam.initialCpbRemovalDelayLengthMinus1
+       RVCE_CS(0x00000017); //hrdParam.cpbRemovalDelayLengthMinus1
+       RVCE_CS(0x00000017); //hrdParam.dpbOutputDelayLengthMinus1
+       RVCE_CS(0x00000018); //hrdParam.timeOffsetLength
+       RVCE_CS(0x00000000); //lowDelayHRDFlag
+       RVCE_CS(0x00000000); //picStructPresentFlag
+       RVCE_CS(0x00000000); //bitstreamRestrictionPresentFlag
+       RVCE_CS(0x00000001); //bitstreamRestrictions.motionVectorsOverPicBoundariesFlag
+       RVCE_CS(0x00000002); //bitstreamRestrictions.maxBytesPerPicDenom
+       RVCE_CS(0x00000001); //bitstreamRestrictions.maxBitsPerMbDenom
+       RVCE_CS(0x00000010); //bitstreamRestrictions.log2MaxMvLengthHori
+       RVCE_CS(0x00000010); //bitstreamRestrictions.log2MaxMvLengthVert
+       RVCE_CS(0x00000003); //bitstreamRestrictions.numReorderFrames
+       RVCE_CS(0x00000003); //bitstreamRestrictions.maxDecFrameBuffering
+       RVCE_END();
+}
+
+static void config(struct rvce_encoder *enc)
+{
+       enc->task_info(enc, 0x00000002, 0, 0xffffffff, 0);
+       enc->rate_control(enc);
+       enc->config_extension(enc);
+       enc->motion_estimation(enc);
+       enc->rdo(enc);
+       if (enc->use_vui)
+               enc->vui(enc);
+       enc->pic_control(enc);
+}
+
 static void encode(struct rvce_encoder *enc)
 {
+       signed luma_offset, chroma_offset;
        int i;
-       unsigned luma_offset, chroma_offset;
 
-       task_info(enc, 0x00000003);
+       enc->task_info(enc, 0x00000003, 0, 0, 0);
 
        RVCE_BEGIN(0x05000001); // context buffer
-       RVCE_READWRITE(enc->cpb.cs_handle, enc->cpb.domain); // encodeContextAddressHi
-       RVCE_CS(0x00000000); // encodeContextAddressLo
+       RVCE_READWRITE(enc->cpb.res->cs_buf, enc->cpb.res->domains, 0x0); // encodeContextAddressHi/Lo
        RVCE_END();
 
        RVCE_BEGIN(0x05000004); // video bitstream buffer
-       RVCE_WRITE(enc->bs_handle, RADEON_DOMAIN_GTT); // videoBitstreamRingAddressHi
-       RVCE_CS(0x00000000); // videoBitstreamRingAddressLo
+       RVCE_WRITE(enc->bs_handle, RADEON_DOMAIN_GTT, 0x0); // videoBitstreamRingAddressHi/Lo
        RVCE_CS(enc->bs_size); // videoBitstreamRingSize
        RVCE_END();
 
@@ -270,10 +319,10 @@ static void encode(struct rvce_encoder *enc)
        RVCE_CS(0x00000000); // insertAUD
        RVCE_CS(0x00000000); // endOfSequence
        RVCE_CS(0x00000000); // endOfStream
-       RVCE_READ(enc->handle, RADEON_DOMAIN_VRAM); // inputPictureLumaAddressHi
-       RVCE_CS(enc->luma->level[0].offset); // inputPictureLumaAddressLo
-       RVCE_READ(enc->handle, RADEON_DOMAIN_VRAM); // inputPictureChromaAddressHi
-       RVCE_CS(enc->chroma->level[0].offset); // inputPictureChromaAddressLo
+       RVCE_READ(enc->handle, RADEON_DOMAIN_VRAM,
+                 enc->luma->level[0].offset); // inputPictureLumaAddressHi/Lo
+       RVCE_READ(enc->handle, RADEON_DOMAIN_VRAM,
+                 enc->chroma->level[0].offset); // inputPictureChromaAddressHi/Lo
        RVCE_CS(align(enc->luma->npix_y, 16)); // encInputFrameYPitch
        RVCE_CS(enc->luma->level[0].pitch_bytes); // encInputPicLumaPitch
        RVCE_CS(enc->chroma->level[0].pitch_bytes); // encInputPicChromaPitch
@@ -283,7 +332,7 @@ static void encode(struct rvce_encoder *enc)
        RVCE_CS(enc->pic.picture_type == PIPE_H264_ENC_PICTURE_TYPE_IDR); // encIdrFlag
        RVCE_CS(0x00000000); // encIdrPicId
        RVCE_CS(0x00000000); // encMGSKeyPic
-       RVCE_CS(0x00000001); // encReferenceFlag
+       RVCE_CS(!enc->pic.not_referenced); // encReferenceFlag
        RVCE_CS(0x00000000); // encTemporalLayerIndex
        RVCE_CS(0x00000000); // num_ref_idx_active_override_flag
        RVCE_CS(0x00000000); // num_ref_idx_l0_active_minus1
@@ -315,7 +364,7 @@ static void encode(struct rvce_encoder *enc)
        if(enc->pic.picture_type == PIPE_H264_ENC_PICTURE_TYPE_P ||
           enc->pic.picture_type == PIPE_H264_ENC_PICTURE_TYPE_B) {
                struct rvce_cpb_slot *l0 = l0_slot(enc);
-               frame_offset(enc, l0, &luma_offset, &chroma_offset);
+               rvce_frame_offset(enc, l0, &luma_offset, &chroma_offset);
                RVCE_CS(l0->picture_type); // encPicType
                RVCE_CS(l0->frame_num); // frameNumber
                RVCE_CS(l0->pic_order_cnt); // pictureOrderCount
@@ -341,7 +390,7 @@ static void encode(struct rvce_encoder *enc)
        RVCE_CS(0x00000000); // pictureStructure
        if(enc->pic.picture_type == PIPE_H264_ENC_PICTURE_TYPE_B) {
                struct rvce_cpb_slot *l1 = l1_slot(enc);
-               frame_offset(enc, l1, &luma_offset, &chroma_offset);
+               rvce_frame_offset(enc, l1, &luma_offset, &chroma_offset);
                RVCE_CS(l1->picture_type); // encPicType
                RVCE_CS(l1->frame_num); // frameNumber
                RVCE_CS(l1->pic_order_cnt); // pictureOrderCount
@@ -355,7 +404,7 @@ static void encode(struct rvce_encoder *enc)
                RVCE_CS(0xffffffff); // chromaOffset
        }
 
-       frame_offset(enc, current_slot(enc), &luma_offset, &chroma_offset);
+       rvce_frame_offset(enc, current_slot(enc), &luma_offset, &chroma_offset);
        RVCE_CS(luma_offset); // encReconstructedLumaOffset
        RVCE_CS(chroma_offset); // encReconstructedChromaOffset
        RVCE_CS(0x00000000); // encColocBufferOffset
@@ -376,7 +425,7 @@ static void encode(struct rvce_encoder *enc)
 
 static void destroy(struct rvce_encoder *enc)
 {
-       task_info(enc, 0x00000001);
+       enc->task_info(enc, 0x00000001, 0, 0, 0);
 
        RVCE_BEGIN(0x02000001); // destroy
        RVCE_END();
@@ -385,6 +434,7 @@ static void destroy(struct rvce_encoder *enc)
 void radeon_vce_40_2_2_init(struct rvce_encoder *enc)
 {
        enc->session = session;
+       enc->task_info = task_info;
        enc->create = create;
        enc->feedback = feedback;
        enc->rate_control = rate_control;
@@ -392,6 +442,8 @@ void radeon_vce_40_2_2_init(struct rvce_encoder *enc)
        enc->pic_control = pic_control;
        enc->motion_estimation = motion_estimation;
        enc->rdo = rdo;
+       enc->vui = vui;
+       enc->config = config;
        enc->encode = encode;
        enc->destroy = destroy;
 }