From c73670990cb12cd92a3f9af1b298204ea781f0dd Mon Sep 17 00:00:00 2001 From: "julien.isorce" Date: Fri, 17 Jul 2015 11:00:07 +0100 Subject: [PATCH] WIP - nouveau: add support for chunk decoding Signed-off-by: Julien Isorce --- src/gallium/drivers/nouveau/nvc0/nvc0_video.c | 67 +++++-- src/gallium/drivers/nouveau/nvc0/nvc0_video.h | 24 +++ src/gallium/drivers/nouveau/nvc0/nvc0_video_bsp.c | 204 ++++++++++++++++++++++ 3 files changed, 285 insertions(+), 10 deletions(-) diff --git a/src/gallium/drivers/nouveau/nvc0/nvc0_video.c b/src/gallium/drivers/nouveau/nvc0/nvc0_video.c index 48ffac1..74fae56 100644 --- a/src/gallium/drivers/nouveau/nvc0/nvc0_video.c +++ b/src/gallium/drivers/nouveau/nvc0/nvc0_video.c @@ -25,6 +25,26 @@ #include "util/u_sampler.h" #include "util/u_format.h" +#include + +//FIXME: hack for now, it should go in struct nouveau_vp3_decoder +static uint32_t mycaps = 0; + +static void +nvc0_decoder_begin_frame(struct pipe_video_codec *decoder, + struct pipe_video_buffer *target, + struct pipe_picture_desc *picture) +{ + struct nouveau_vp3_decoder *dec = (struct nouveau_vp3_decoder *)decoder; + + assert(decoder); + + uint32_t comm_seq = ++dec->fence_seq; + mycaps = 0; + + printf("-- begin comm_seq: %u \n", comm_seq); +} + static void nvc0_decoder_decode_bitstream(struct pipe_video_codec *decoder, struct pipe_video_buffer *video_target, @@ -35,25 +55,50 @@ nvc0_decoder_decode_bitstream(struct pipe_video_codec *decoder, { struct nouveau_vp3_decoder *dec = (struct nouveau_vp3_decoder *)decoder; struct nouveau_vp3_video_buffer *target = (struct nouveau_vp3_video_buffer *)video_target; - uint32_t comm_seq = ++dec->fence_seq; + uint32_t comm_seq = dec->fence_seq; union pipe_desc desc; + unsigned ret = 0; - unsigned vp_caps, is_ref, ret; - struct nouveau_vp3_video_buffer *refs[16] = {}; + assert(dec); + assert(target); + assert(target->base.buffer_format == PIPE_FORMAT_NV12); desc.base = picture; - assert(target->base.buffer_format == PIPE_FORMAT_NV12); - - ret = nvc0_decoder_bsp(dec, desc, target, comm_seq, - num_buffers, data, num_bytes, - &vp_caps, &is_ref, refs); + ret = nvc0_decoder_bsp_next_chunk(dec, desc, target, comm_seq, + num_buffers, data, num_bytes, + &mycaps); /* did we decode bitstream correctly? */ assert(ret == 2); - nvc0_decoder_vp(dec, desc, target, comm_seq, vp_caps, is_ref, refs); - nvc0_decoder_ppp(dec, desc, target, comm_seq); + printf("-- decode ret: %u, nv12: %d\n", ret, target->base.buffer_format == PIPE_FORMAT_NV12); +} + +static void +nvc0_decoder_end_frame(struct pipe_video_codec *decoder, + struct pipe_video_buffer *video_target, + struct pipe_picture_desc *picture) +{ + struct nouveau_vp3_decoder *dec = (struct nouveau_vp3_decoder *)decoder; + struct nouveau_vp3_video_buffer *target = (struct nouveau_vp3_video_buffer *)video_target; + uint32_t comm_seq = dec->fence_seq; + union pipe_desc desc; + unsigned vp_caps = 0; + unsigned is_ref = 0; + + struct nouveau_vp3_video_buffer *refs[16] = {}; + + assert(dec); + + desc.base = picture; + + nvc0_decoder_bsp_end_frame(dec, desc, target, comm_seq, mycaps, &vp_caps, &is_ref, refs); + + nvc0_decoder_vp(dec, desc, target, comm_seq, vp_caps, is_ref, refs); + nvc0_decoder_ppp(dec, desc, target, comm_seq); + + printf("-- /end comm_seq: %u \n", comm_seq); } struct pipe_video_codec * @@ -164,7 +209,9 @@ nvc0_create_decoder(struct pipe_context *context, PUSH_DATA (push[2], dec->ppp->handle); dec->base.context = context; + dec->base.begin_frame = nvc0_decoder_begin_frame; dec->base.decode_bitstream = nvc0_decoder_decode_bitstream; + dec->base.end_frame = nvc0_decoder_end_frame; for (i = 0; i < NOUVEAU_VP3_VIDEO_QDEPTH && !ret; ++i) ret = nouveau_bo_new(screen->device, NOUVEAU_BO_VRAM, diff --git a/src/gallium/drivers/nouveau/nvc0/nvc0_video.h b/src/gallium/drivers/nouveau/nvc0/nvc0_video.h index 9ee0280..39f73c8 100644 --- a/src/gallium/drivers/nouveau/nvc0/nvc0_video.h +++ b/src/gallium/drivers/nouveau/nvc0/nvc0_video.h @@ -29,6 +29,30 @@ #include "util/u_video.h" + +extern unsigned +nvc0_decoder_bsp_next_chunk(struct nouveau_vp3_decoder *dec, union pipe_desc desc, + struct nouveau_vp3_video_buffer *target, + unsigned comm_seq, unsigned num_buffers, + const void *const *data, const unsigned *num_bytes, uint32_t *caps); + +extern unsigned +nvc0_decoder_bsp_end_frame(struct nouveau_vp3_decoder *dec, union pipe_desc desc, + struct nouveau_vp3_video_buffer *target, + unsigned comm_seq, + uint32_t caps, unsigned *vp_caps, unsigned *is_ref, + struct nouveau_vp3_video_buffer *refs[16]); + +//TODO: rename to nvc0_decoder_bsp +extern unsigned +nvc0_decoder_bsp2(struct nouveau_vp3_decoder *dec, union pipe_desc desc, + struct nouveau_vp3_video_buffer *target, + unsigned comm_seq, unsigned num_buffers, + const void *const *data, const unsigned *num_bytes, + unsigned *vp_caps, unsigned *is_ref, + struct nouveau_vp3_video_buffer *refs[16]); + +//TODO: remove this function it will be replaced by nvc0_decoder_bsp2 extern unsigned nvc0_decoder_bsp(struct nouveau_vp3_decoder *dec, union pipe_desc desc, struct nouveau_vp3_video_buffer *target, diff --git a/src/gallium/drivers/nouveau/nvc0/nvc0_video_bsp.c b/src/gallium/drivers/nouveau/nvc0/nvc0_video_bsp.c index 9139bc1..23180c5 100644 --- a/src/gallium/drivers/nouveau/nvc0/nvc0_video_bsp.c +++ b/src/gallium/drivers/nouveau/nvc0/nvc0_video_bsp.c @@ -31,6 +31,210 @@ static void dump_comm_bsp(struct comm *comm) } #endif + +unsigned +nvc0_decoder_bsp_next_chunk(struct nouveau_vp3_decoder *dec, union pipe_desc desc, + struct nouveau_vp3_video_buffer *target, + unsigned comm_seq, unsigned num_buffers, + const void *const *data, const unsigned *num_bytes, uint32_t *caps) +{ + struct nouveau_bo *bsp_bo = dec->bsp_bo[comm_seq % NOUVEAU_VP3_VIDEO_QDEPTH]; + struct nouveau_bo *inter_bo = dec->inter_bo[comm_seq & 1]; + uint32_t bsp_size = 0; + uint32_t i = 0; + unsigned ret = 0; + + assert(caps); + + bsp_size = NOUVEAU_VP3_BSP_RESERVED_SIZE; + for (i = 0; i < num_buffers; i++) + bsp_size += num_bytes[i]; + bsp_size += 256; /* the 4 end markers */ + + if (!bsp_bo || bsp_size > bsp_bo->size) { + union nouveau_bo_config cfg; + struct nouveau_bo *tmp_bo = NULL; + + cfg.nvc0.tile_mode = 0x10; + cfg.nvc0.memtype = 0xfe; + + /* round up to the nearest mb */ + bsp_size += (1 << 20) - 1; + bsp_size &= ~((1 << 20) - 1); + + ret = nouveau_bo_new(dec->bitplane_bo->device, NOUVEAU_BO_VRAM, 0, bsp_size, &cfg, &tmp_bo); + if (ret) { + debug_printf("reallocating bsp %u -> %u failed with %i\n", + bsp_bo ? (unsigned)bsp_bo->size : 0, bsp_size, ret); + return -1; + } + nouveau_bo_ref(NULL, &bsp_bo); + dec->bsp_bo[comm_seq % NOUVEAU_VP3_VIDEO_QDEPTH] = bsp_bo = tmp_bo; + } + + if (!inter_bo || bsp_bo->size * 4 > inter_bo->size) { + union nouveau_bo_config cfg; + struct nouveau_bo *tmp_bo = NULL; + + cfg.nvc0.tile_mode = 0x10; + cfg.nvc0.memtype = 0xfe; + + ret = nouveau_bo_new(dec->bitplane_bo->device, NOUVEAU_BO_VRAM, 0, bsp_bo->size * 4, &cfg, &tmp_bo); + if (ret) { + debug_printf("reallocating inter %u -> %u failed with %i\n", + inter_bo ? (unsigned)inter_bo->size : 0, (unsigned)bsp_bo->size * 4, ret); + return -1; + } + nouveau_bo_ref(NULL, &inter_bo); + dec->inter_bo[comm_seq & 1] = inter_bo = tmp_bo; + } + + ret = nouveau_bo_map(bsp_bo, NOUVEAU_BO_WR, dec->client); + if (ret) { + debug_printf("map failed: %i %s\n", ret, strerror(-ret)); + return -1; + } + + *caps = nouveau_vp3_bsp(dec, desc, target, comm_seq, + num_buffers, data, num_bytes); + + return 2; +} + + +unsigned +nvc0_decoder_bsp_end_frame(struct nouveau_vp3_decoder *dec, union pipe_desc desc, + struct nouveau_vp3_video_buffer *target, + unsigned comm_seq, + uint32_t caps, unsigned *vp_caps, unsigned *is_ref, + struct nouveau_vp3_video_buffer *refs[16]) +{ + struct nouveau_pushbuf *push = dec->pushbuf[0]; + enum pipe_video_format codec = u_reduce_video_profile(dec->base.profile); + uint32_t bsp_addr, comm_addr, inter_addr; + uint32_t slice_size, bucket_size, ring_size; + struct nouveau_bo *bsp_bo = dec->bsp_bo[comm_seq % NOUVEAU_VP3_VIDEO_QDEPTH]; + struct nouveau_bo *inter_bo = dec->inter_bo[comm_seq & 1]; + unsigned fence_extra = 0; + struct nouveau_pushbuf_refn bo_refs[] = { + { bsp_bo, NOUVEAU_BO_RD | NOUVEAU_BO_VRAM }, + { inter_bo, NOUVEAU_BO_WR | NOUVEAU_BO_VRAM }, + #if NOUVEAU_VP3_DEBUG_FENCE + { dec->fence_bo, NOUVEAU_BO_WR | NOUVEAU_BO_GART }, + #endif + { dec->bitplane_bo, NOUVEAU_BO_RDWR | NOUVEAU_BO_VRAM }, + }; + int num_refs = sizeof(bo_refs)/sizeof(*bo_refs); + + if (!dec->bitplane_bo) + num_refs--; + + #if NOUVEAU_VP3_DEBUG_FENCE + fence_extra = 4; + #endif + + bo_refs[0].bo = dec->bsp_bo[comm_seq % NOUVEAU_VP3_VIDEO_QDEPTH]; + bo_refs[1].bo = dec->inter_bo[comm_seq & 1]; + + nouveau_vp3_vp_caps(dec, desc, target, comm_seq, vp_caps, is_ref, refs); + + nouveau_pushbuf_space(push, 6 + (codec == PIPE_VIDEO_FORMAT_MPEG4_AVC ? 9 : 7) + fence_extra + 2, num_refs, 0); + nouveau_pushbuf_refn(push, bo_refs, num_refs); + + bsp_addr = bsp_bo->offset >> 8; + inter_addr = inter_bo->offset >> 8; + + #if NOUVEAU_VP3_DEBUG_FENCE + memset(dec->comm, 0, 0x200); + comm_addr = (dec->fence_bo->offset + COMM_OFFSET) >> 8; + #else + comm_addr = bsp_addr + (COMM_OFFSET>>8); + #endif + + BEGIN_NVC0(push, SUBC_BSP(0x700), 5); + PUSH_DATA (push, caps); // 700 cmd + PUSH_DATA (push, bsp_addr + 1); // 704 strparm_bsp + PUSH_DATA (push, bsp_addr + 7); // 708 str addr + PUSH_DATA (push, comm_addr); // 70c comm + PUSH_DATA (push, comm_seq); // 710 seq + + if (codec != PIPE_VIDEO_FORMAT_MPEG4_AVC) { + u32 bitplane_addr; + + bitplane_addr = dec->bitplane_bo->offset >> 8; + + nouveau_vp3_inter_sizes(dec, 1, &slice_size, &bucket_size, &ring_size); + BEGIN_NVC0(push, SUBC_BSP(0x400), 6); + PUSH_DATA (push, bsp_addr); // 400 picparm addr + PUSH_DATA (push, inter_addr); // 404 interparm addr + PUSH_DATA (push, inter_addr + slice_size + bucket_size); // 408 interdata addr + PUSH_DATA (push, ring_size << 8); // 40c interdata_size + PUSH_DATA (push, bitplane_addr); // 410 BITPLANE_DATA + PUSH_DATA (push, 0x400); // 414 BITPLANE_DATA_SIZE + } else { + nouveau_vp3_inter_sizes(dec, desc.h264->slice_count, &slice_size, &bucket_size, &ring_size); + BEGIN_NVC0(push, SUBC_BSP(0x400), 8); + PUSH_DATA (push, bsp_addr); // 400 picparm addr + PUSH_DATA (push, inter_addr); // 404 interparm addr + PUSH_DATA (push, slice_size << 8); // 408 interparm size? + PUSH_DATA (push, inter_addr + slice_size + bucket_size); // 40c interdata addr + PUSH_DATA (push, ring_size << 8); // 410 interdata size + PUSH_DATA (push, inter_addr + slice_size); // 414 bucket? + PUSH_DATA (push, bucket_size << 8); // 418 bucket size? unshifted.. + PUSH_DATA (push, 0); // 41c targets + // TODO: Double check 414 / 418 with nvidia trace + } + + #if NOUVEAU_VP3_DEBUG_FENCE + BEGIN_NVC0(push, SUBC_BSP(0x240), 3); + PUSH_DATAh(push, dec->fence_bo->offset); + PUSH_DATA (push, dec->fence_bo->offset); + PUSH_DATA (push, dec->fence_seq); + + BEGIN_NVC0(push, SUBC_BSP(0x300), 1); + PUSH_DATA (push, 1); + PUSH_KICK (push); + + { + unsigned spin = 0; + do { + usleep(100); + if ((spin++ & 0xff) == 0xff) { + debug_printf("b%u: %u\n", dec->fence_seq, dec->fence_map[0]); + dump_comm_bsp(dec->comm); + } + } while (dec->fence_seq > dec->fence_map[0]); + } + + dump_comm_bsp(dec->comm); + return dec->comm->status[comm_seq & 0xf]; + #else + BEGIN_NVC0(push, SUBC_BSP(0x300), 1); + PUSH_DATA (push, 0); + PUSH_KICK (push); + return 2; + #endif +} + +//TODO: rename to nvc0_decoder_bsp +unsigned +nvc0_decoder_bsp2(struct nouveau_vp3_decoder *dec, union pipe_desc desc, + struct nouveau_vp3_video_buffer *target, + unsigned comm_seq, unsigned num_buffers, + const void *const *data, const unsigned *num_bytes, + unsigned *vp_caps, unsigned *is_ref, + struct nouveau_vp3_video_buffer *refs[16]) +{ + uint32_t caps = 0; + unsigned ret = nvc0_decoder_bsp_next_chunk(dec, desc, target, comm_seq, num_buffers, data, num_bytes, &caps); + + if (ret != 2) + return ret; + + return nvc0_decoder_bsp_end_frame(dec, desc, target, comm_seq, caps, vp_caps, is_ref, refs); +} + +//FIXME: remove this function unsigned nvc0_decoder_bsp(struct nouveau_vp3_decoder *dec, union pipe_desc desc, struct nouveau_vp3_video_buffer *target, -- 2.3.2 (Apple Git-55)