freedreno/a4xx: add s8/z32/z32_s8x24 support

Signed-off-by: Rob Clark <robclark@freedesktop.org>

diff --git a/src/gallium/drivers/freedreno/a4xx/fd4_emit.c b/src/gallium/drivers/freedreno/a4xx/fd4_emit.c
index e7c210d..c3226d5 100644 (file)
--- a/src/gallium/drivers/freedreno/a4xx/fd4_emit.c
+++ b/src/gallium/drivers/freedreno/a4xx/fd4_emit.c
@@ -200,23 +200,13 @@ void
 fd4_emit_gmem_restore_tex(struct fd_ringbuffer *ring, unsigned nr_bufs,
                struct pipe_surface **bufs)
 {
-       unsigned char mrt_comp[A4XX_MAX_RENDER_TARGETS] = {0};
+       unsigned char mrt_comp[A4XX_MAX_RENDER_TARGETS];
        int i;
 
        for (i = 0; i < A4XX_MAX_RENDER_TARGETS; i++) {
                mrt_comp[i] = (i < nr_bufs) ? 0xf : 0;
        }
 
-       OUT_PKT0(ring, REG_A4XX_RB_RENDER_COMPONENTS, 1);
-       OUT_RING(ring, A4XX_RB_RENDER_COMPONENTS_RT0(mrt_comp[0]) |
-                       A4XX_RB_RENDER_COMPONENTS_RT1(mrt_comp[1]) |
-                       A4XX_RB_RENDER_COMPONENTS_RT2(mrt_comp[2]) |
-                       A4XX_RB_RENDER_COMPONENTS_RT3(mrt_comp[3]) |
-                       A4XX_RB_RENDER_COMPONENTS_RT4(mrt_comp[4]) |
-                       A4XX_RB_RENDER_COMPONENTS_RT5(mrt_comp[5]) |
-                       A4XX_RB_RENDER_COMPONENTS_RT6(mrt_comp[6]) |
-                       A4XX_RB_RENDER_COMPONENTS_RT7(mrt_comp[7]));
-
        /* output sampler state: */
        OUT_PKT3(ring, CP_LOAD_STATE, 2 + (2 * nr_bufs));
        OUT_RING(ring, CP_LOAD_STATE_0_DST_OFF(0) |
@@ -250,6 +240,25 @@ fd4_emit_gmem_restore_tex(struct fd_ringbuffer *ring, unsigned nr_bufs,
                        uint32_t offset = fd_resource_offset(rsc, lvl, bufs[i]->u.tex.first_layer);
                        enum pipe_format format = fd4_gmem_restore_format(bufs[i]->format);
 
+                       /* The restore blit_zs shader expects stencil in sampler 0,
+                        * and depth in sampler 1
+                        */
+                       if (rsc->stencil && (i == 0)) {
+                               rsc = rsc->stencil;
+                               format = fd4_gmem_restore_format(rsc->base.b.format);
+                       }
+
+                       /* z32 restore is accomplished using depth write.  If there is
+                        * no stencil component (ie. PIPE_FORMAT_Z32_FLOAT_S8X24_UINT)
+                        * then no render target:
+                        *
+                        * (The same applies for z32_s8x24, since for stencil sampler
+                        * state the above 'if' will replace 'format' with s8)
+                        */
+                       if ((format == PIPE_FORMAT_Z32_FLOAT) ||
+                                       (format == PIPE_FORMAT_Z32_FLOAT_S8X24_UINT))
+                               mrt_comp[i] = 0;
+
                        debug_assert(bufs[i]->u.tex.first_layer == bufs[i]->u.tex.last_layer);
 
                        OUT_RING(ring, A4XX_TEX_CONST_0_FMT(fd4_pipe2tex(format)) |
@@ -281,6 +290,16 @@ fd4_emit_gmem_restore_tex(struct fd_ringbuffer *ring, unsigned nr_bufs,
                        OUT_RING(ring, 0x00000000);
                }
        }
+
+       OUT_PKT0(ring, REG_A4XX_RB_RENDER_COMPONENTS, 1);
+       OUT_RING(ring, A4XX_RB_RENDER_COMPONENTS_RT0(mrt_comp[0]) |
+                       A4XX_RB_RENDER_COMPONENTS_RT1(mrt_comp[1]) |
+                       A4XX_RB_RENDER_COMPONENTS_RT2(mrt_comp[2]) |
+                       A4XX_RB_RENDER_COMPONENTS_RT3(mrt_comp[3]) |
+                       A4XX_RB_RENDER_COMPONENTS_RT4(mrt_comp[4]) |
+                       A4XX_RB_RENDER_COMPONENTS_RT5(mrt_comp[5]) |
+                       A4XX_RB_RENDER_COMPONENTS_RT6(mrt_comp[6]) |
+                       A4XX_RB_RENDER_COMPONENTS_RT7(mrt_comp[7]));
 }
 
 void

diff --git a/src/gallium/drivers/freedreno/a4xx/fd4_format.c b/src/gallium/drivers/freedreno/a4xx/fd4_format.c
index 29abe0b..f906a67 100644 (file)
--- a/src/gallium/drivers/freedreno/a4xx/fd4_format.c
+++ b/src/gallium/drivers/freedreno/a4xx/fd4_format.c
@@ -89,6 +89,14 @@ static struct fd4_format formats[PIPE_FORMAT_COUNT] = {
        _T(L8_UNORM,   8_UNORM, R8_UNORM, WZYX),
        _T(I8_UNORM,   8_UNORM, NONE,     WZYX),
 
+       /* NOTE: should be TFMT_8_UINT (which then gets remapped to
+        * TFMT_8_UNORM for mem2gmem in _gmem_restore_format()), but
+        * we don't know TFMT_8_UINT yet.. so just use TFMT_8_UNORM
+        * for now.. sampling from stencil as a texture might not
+        * work right, but at least should be fine for zsbuf..
+        */
+       _T(S8_UINT,    8_UNORM,  R8_UNORM, WZYX),
+
        /* 16-bit */
        V_(R16_UNORM,   16_UNORM, NONE,     WZYX),
        V_(R16_SNORM,   16_SNORM, NONE,     WZYX),
@@ -191,7 +199,8 @@ static struct fd4_format formats[PIPE_FORMAT_COUNT] = {
 
        _T(Z24X8_UNORM,       X8Z24_UNORM, R8G8B8A8_UNORM, WZYX),
        _T(Z24_UNORM_S8_UINT, X8Z24_UNORM, R8G8B8A8_UNORM, WZYX),
-       /*_T(Z32_FLOAT,         Z32_FLOAT,   R8G8B8A8_UNORM, WZYX),*/
+       _T(Z32_FLOAT,         32_FLOAT,   R8G8B8A8_UNORM, WZYX),
+       _T(Z32_FLOAT_S8X24_UINT, 32_FLOAT,R8G8B8A8_UNORM, WZYX),
 
        /* 48-bit */
        V_(R16G16B16_UNORM,   16_16_16_UNORM, NONE, WZYX),
@@ -282,6 +291,9 @@ fd4_pipe2swap(enum pipe_format format)
 enum a4xx_tex_fetchsize
 fd4_pipe2fetchsize(enum pipe_format format)
 {
+       if (format == PIPE_FORMAT_Z32_FLOAT_S8X24_UINT)
+               format = PIPE_FORMAT_Z32_FLOAT;
+
        switch (util_format_get_blocksizebits(format)) {
        case 8:   return TFETCH4_1_BYTE;
        case 16:  return TFETCH4_2_BYTE;
@@ -312,6 +324,8 @@ fd4_gmem_restore_format(enum pipe_format format)
                return PIPE_FORMAT_R8G8B8A8_UNORM;
        case PIPE_FORMAT_Z16_UNORM:
                return PIPE_FORMAT_R8G8_UNORM;
+       case PIPE_FORMAT_S8_UINT:
+               return PIPE_FORMAT_R8_UNORM;
        default:
                return format;
        }
@@ -328,6 +342,9 @@ fd4_pipe2depth(enum pipe_format format)
        case PIPE_FORMAT_X8Z24_UNORM:
        case PIPE_FORMAT_S8_UINT_Z24_UNORM:
                return DEPTH4_24_8;
+       case PIPE_FORMAT_Z32_FLOAT:
+       case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
+               return DEPTH4_32;
        default:
                return ~0;
        }

diff --git a/src/gallium/drivers/freedreno/a4xx/fd4_gmem.c b/src/gallium/drivers/freedreno/a4xx/fd4_gmem.c
index 890e71e..6541da5 100644 (file)
--- a/src/gallium/drivers/freedreno/a4xx/fd4_gmem.c
+++ b/src/gallium/drivers/freedreno/a4xx/fd4_gmem.c
@@ -68,11 +68,23 @@ emit_mrt(struct fd_ringbuffer *ring, unsigned nr_bufs,
 
                if ((i < nr_bufs) && bufs[i]) {
                        struct pipe_surface *psurf = bufs[i];
+                       enum pipe_format pformat = 0;
 
                        rsc = fd_resource(psurf->texture);
+                       pformat = psurf->format;
+
+                       /* In case we're drawing to Z32F_S8, the "color" actually goes to
+                        * the stencil
+                        */
+                       if (rsc->stencil) {
+                               rsc = rsc->stencil;
+                               pformat = rsc->base.b.format;
+                               bases++;
+                       }
+
                        slice = fd_resource_slice(rsc, psurf->u.tex.level);
-                       format = fd4_pipe2color(psurf->format);
-                       swap = fd4_pipe2swap(psurf->format);
+                       format = fd4_pipe2color(pformat);
+                       swap = fd4_pipe2swap(pformat);
 
                        debug_assert(psurf->u.tex.first_layer == psurf->u.tex.last_layer);
 
@@ -114,13 +126,23 @@ emit_mrt(struct fd_ringbuffer *ring, unsigned nr_bufs,
 /* transfer from gmem to system memory (ie. normal RAM) */
 
 static void
-emit_gmem2mem_surf(struct fd_context *ctx,
+emit_gmem2mem_surf(struct fd_context *ctx, bool stencil,
                uint32_t base, struct pipe_surface *psurf)
 {
        struct fd_ringbuffer *ring = ctx->ring;
        struct fd_resource *rsc = fd_resource(psurf->texture);
-       struct fd_resource_slice *slice = &rsc->slices[psurf->u.tex.level];
-       uint32_t offset = fd_resource_offset(rsc, psurf->u.tex.level,
+       enum pipe_format pformat = psurf->format;
+       struct fd_resource_slice *slice;
+       uint32_t offset;
+
+       if (stencil) {
+               debug_assert(rsc->stencil);
+               rsc = rsc->stencil;
+               pformat = rsc->base.b.format;
+       }
+
+       slice = &rsc->slices[psurf->u.tex.level];
+       offset = fd_resource_offset(rsc, psurf->u.tex.level,
                        psurf->u.tex.first_layer);
 
        debug_assert(psurf->u.tex.first_layer == psurf->u.tex.last_layer);
@@ -132,10 +154,10 @@ emit_gmem2mem_surf(struct fd_context *ctx,
        OUT_RELOCW(ring, rsc->bo, offset, 0, 0);   /* RB_COPY_DEST_BASE */
        OUT_RING(ring, A4XX_RB_COPY_DEST_PITCH_PITCH(slice->pitch * rsc->cpp));
        OUT_RING(ring, A4XX_RB_COPY_DEST_INFO_TILE(TILE4_LINEAR) |
-                       A4XX_RB_COPY_DEST_INFO_FORMAT(fd4_pipe2color(psurf->format)) |
+                       A4XX_RB_COPY_DEST_INFO_FORMAT(fd4_pipe2color(pformat)) |
                        A4XX_RB_COPY_DEST_INFO_COMPONENT_ENABLE(0xf) |
                        A4XX_RB_COPY_DEST_INFO_ENDIAN(ENDIAN_NONE) |
-                       A4XX_RB_COPY_DEST_INFO_SWAP(fd4_pipe2swap(psurf->format)));
+                       A4XX_RB_COPY_DEST_INFO_SWAP(fd4_pipe2swap(pformat)));
 
        fd4_draw(ctx, ring, DI_PT_RECTLIST, IGNORE_VISIBILITY,
                        DI_SRC_SEL_AUTO_INDEX, 2, 1, INDEX_SIZE_IGN, 0, 0, NULL);
@@ -226,7 +248,11 @@ fd4_emit_tile_gmem2mem(struct fd_context *ctx, struct fd_tile *tile)
        fd4_emit_vertex_bufs(ring, &emit);
 
        if (ctx->resolve & (FD_BUFFER_DEPTH | FD_BUFFER_STENCIL)) {
-               emit_gmem2mem_surf(ctx, gmem->zsbuf_base[0], pfb->zsbuf);
+               struct fd_resource *rsc = fd_resource(pfb->zsbuf->texture);
+               if (!rsc->stencil || (ctx->resolve & FD_BUFFER_DEPTH))
+                       emit_gmem2mem_surf(ctx, false, ctx->gmem.zsbuf_base[0], pfb->zsbuf);
+               if (rsc->stencil && (ctx->resolve & FD_BUFFER_STENCIL))
+                       emit_gmem2mem_surf(ctx, true, ctx->gmem.zsbuf_base[1], pfb->zsbuf);
        }
 
        if (ctx->resolve & FD_BUFFER_COLOR) {
@@ -236,7 +262,7 @@ fd4_emit_tile_gmem2mem(struct fd_context *ctx, struct fd_tile *tile)
                                continue;
                        if (!(ctx->resolve & (PIPE_CLEAR_COLOR0 << i)))
                                continue;
-                       emit_gmem2mem_surf(ctx, gmem->cbuf_base[i], pfb->cbufs[i]);
+                       emit_gmem2mem_surf(ctx, false, gmem->cbuf_base[i], pfb->cbufs[i]);
                }
        }
 
@@ -254,9 +280,20 @@ emit_mem2gmem_surf(struct fd_context *ctx, uint32_t *bases,
                struct pipe_surface **bufs, uint32_t nr_bufs, uint32_t bin_w)
 {
        struct fd_ringbuffer *ring = ctx->ring;
+       struct pipe_surface *zsbufs[2];
 
        emit_mrt(ring, nr_bufs, bufs, bases, bin_w);
 
+       if (bufs[0] && (bufs[0]->format == PIPE_FORMAT_Z32_FLOAT_S8X24_UINT)) {
+               /* The gmem_restore_tex logic will put the first buffer's stencil
+                * as color. Supply it with the proper information to make that
+                * happen.
+                */
+               zsbufs[0] = zsbufs[1] = bufs[0];
+               bufs = zsbufs;
+               nr_bufs = 2;
+       }
+
        fd4_emit_gmem_restore_tex(ring, nr_bufs, bufs);
 
        fd4_draw(ctx, ring, DI_PT_RECTLIST, IGNORE_VISIBILITY,
@@ -394,13 +431,6 @@ fd4_emit_tile_mem2gmem(struct fd_context *ctx, struct fd_tile *tile)
        bin_w = gmem->bin_w;
        bin_h = gmem->bin_h;
 
-       if (fd_gmem_needs_restore(ctx, tile, FD_BUFFER_DEPTH | FD_BUFFER_STENCIL)) {
-               emit.prog = &ctx->blit_prog[0];
-               emit.fp = NULL;      /* frag shader changed so clear cache */
-               fd4_program_emit(ring, &emit, 1, &pfb->zsbuf);
-               emit_mem2gmem_surf(ctx, gmem->zsbuf_base, &pfb->zsbuf, 1, bin_w);
-       }
-
        if (fd_gmem_needs_restore(ctx, tile, FD_BUFFER_COLOR)) {
                emit.prog = &ctx->blit_prog[pfb->nr_cbufs - 1];
                emit.fp = NULL;      /* frag shader changed so clear cache */
@@ -408,6 +438,40 @@ fd4_emit_tile_mem2gmem(struct fd_context *ctx, struct fd_tile *tile)
                emit_mem2gmem_surf(ctx, gmem->cbuf_base, pfb->cbufs, pfb->nr_cbufs, bin_w);
        }
 
+       if (fd_gmem_needs_restore(ctx, tile, FD_BUFFER_DEPTH | FD_BUFFER_STENCIL)) {
+               switch (pfb->zsbuf->format) {
+               case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
+               case PIPE_FORMAT_Z32_FLOAT:
+                       emit.prog = (pfb->zsbuf->format == PIPE_FORMAT_Z32_FLOAT) ?
+                                       &ctx->blit_z : &ctx->blit_zs;
+                       emit.key.half_precision = false;
+
+                       OUT_PKT0(ring, REG_A4XX_RB_DEPTH_CONTROL, 1);
+                       OUT_RING(ring, A4XX_RB_DEPTH_CONTROL_Z_ENABLE |
+                                       A4XX_RB_DEPTH_CONTROL_Z_WRITE_ENABLE |
+                                       A4XX_RB_DEPTH_CONTROL_ZFUNC(FUNC_ALWAYS) |
+                                       A4XX_RB_DEPTH_CONTROL_EARLY_Z_DISABLE);
+
+                       OUT_PKT0(ring, REG_A4XX_GRAS_ALPHA_CONTROL, 1);
+                       OUT_RING(ring, A4XX_GRAS_ALPHA_CONTROL_ALPHA_TEST_ENABLE);
+
+                       OUT_PKT0(ring, REG_A4XX_GRAS_CL_CLIP_CNTL, 1);
+                       OUT_RING(ring, 0x80000);   /* GRAS_CL_CLIP_CNTL */
+
+                       break;
+               default:
+                       /* Non-float can use a regular color write. It's split over 8-bit
+                        * components, so half precision is always sufficient.
+                        */
+                       emit.prog = &ctx->blit_prog[0];
+                       emit.key.half_precision = true;
+                       break;
+               }
+               emit.fp = NULL;      /* frag shader changed so clear cache */
+               fd4_program_emit(ring, &emit, 1, &pfb->zsbuf);
+               emit_mem2gmem_surf(ctx, gmem->zsbuf_base, &pfb->zsbuf, 1, bin_w);
+       }
+
        OUT_PKT0(ring, REG_A4XX_GRAS_SC_CONTROL, 1);
        OUT_RING(ring, A4XX_GRAS_SC_CONTROL_RENDER_MODE(RB_RENDERING_PASS) |
                        A4XX_GRAS_SC_CONTROL_MSAA_SAMPLES(MSAA_ONE) |
@@ -546,21 +610,35 @@ fd4_emit_tile_prep(struct fd_context *ctx, struct fd_tile *tile)
        struct fd_ringbuffer *ring = ctx->ring;
        struct pipe_framebuffer_state *pfb = &ctx->framebuffer;
        struct fd_gmem_stateobj *gmem = &ctx->gmem;
-       uint32_t reg;
 
-       OUT_PKT0(ring, REG_A4XX_RB_DEPTH_INFO, 3);
-       reg = A4XX_RB_DEPTH_INFO_DEPTH_BASE(gmem->zsbuf_base[0]);
        if (pfb->zsbuf) {
-               reg |= A4XX_RB_DEPTH_INFO_DEPTH_FORMAT(fd4_pipe2depth(pfb->zsbuf->format));
-       }
-       OUT_RING(ring, reg);
-       if (pfb->zsbuf) {
-               uint32_t cpp = util_format_get_blocksize(pfb->zsbuf->format);
+               struct fd_resource *rsc = fd_resource(pfb->zsbuf->texture);
+               uint32_t cpp = rsc->cpp;
+
+               OUT_PKT0(ring, REG_A4XX_RB_DEPTH_INFO, 3);
+               OUT_RING(ring, A4XX_RB_DEPTH_INFO_DEPTH_BASE(gmem->zsbuf_base[0]) |
+                               A4XX_RB_DEPTH_INFO_DEPTH_FORMAT(fd4_pipe2depth(pfb->zsbuf->format)));
                OUT_RING(ring, A4XX_RB_DEPTH_PITCH(cpp * gmem->bin_w));
                OUT_RING(ring, A4XX_RB_DEPTH_PITCH2(cpp * gmem->bin_w));
+
+               OUT_PKT0(ring, REG_A4XX_RB_STENCIL_INFO, 2);
+               if (rsc->stencil) {
+                       OUT_RING(ring, A4XX_RB_STENCIL_INFO_SEPARATE_STENCIL |
+                                       A4XX_RB_STENCIL_INFO_STENCIL_BASE(gmem->zsbuf_base[1]));
+                       OUT_RING(ring, A4XX_RB_STENCIL_PITCH(rsc->stencil->cpp * gmem->bin_w));
+               } else {
+                       OUT_RING(ring, 0x00000000);
+                       OUT_RING(ring, 0x00000000);
+               }
        } else {
+               OUT_PKT0(ring, REG_A4XX_RB_DEPTH_INFO, 3);
+               OUT_RING(ring, 0x00000000);
                OUT_RING(ring, 0x00000000);
                OUT_RING(ring, 0x00000000);
+
+               OUT_PKT0(ring, REG_A4XX_RB_STENCIL_INFO, 2);
+               OUT_RING(ring, 0);            /* RB_STENCIL_INFO */
+               OUT_RING(ring, 0);            /* RB_STENCIL_PITCH */
        }
 
        OUT_PKT0(ring, REG_A4XX_GRAS_DEPTH_CONTROL, 1);

diff --git a/src/gallium/drivers/freedreno/freedreno_util.h b/src/gallium/drivers/freedreno/freedreno_util.h
index b5c5db9..d6a08b5 100644 (file)
--- a/src/gallium/drivers/freedreno/freedreno_util.h
+++ b/src/gallium/drivers/freedreno/freedreno_util.h
@@ -57,7 +57,7 @@ enum adreno_stencil_op fd_stencil_op(unsigned op);
 #define A2XX_MAX_RENDER_TARGETS 1
 #define A3XX_MAX_RENDER_TARGETS 4
 #define A4XX_MAX_RENDER_TARGETS 8
-/* for now until a4xx MRT support: */
+
 #define MAX_RENDER_TARGETS A4XX_MAX_RENDER_TARGETS
 
 #define FD_DBG_MSGS     0x0001