fd5_blitter.c

/*
 * Copyright (C) 2017 Rob Clark <[email protected]>
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 * and/or sell copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice (including the next
 * paragraph) shall be included in all copies or substantial portions of the
 * Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 *
 * Authors:
 *    Rob Clark <[email protected]>
 */

#include "freedreno_blitter.h"
#include "freedreno_resource.h"

#include "fd5_blitter.h"
#include "fd5_emit.h"
#include "fd5_format.h"

/* Make sure none of the requested dimensions extend beyond the size of the
 * resource.  Not entirely sure why this happens, but sometimes it does, and
 * w/ 2d blt doesn't have wrap modes like a sampler, so force those cases
 * back to u_blitter
 */
static bool
ok_dims(const struct pipe_resource *r, const struct pipe_box *b, int lvl)
{
   return (b->x >= 0) && (b->x + b->width <= u_minify(r->width0, lvl)) &&
          (b->y >= 0) && (b->y + b->height <= u_minify(r->height0, lvl)) &&
          (b->z >= 0) && (b->z + b->depth <= u_minify(r->depth0, lvl));
}

/* Not sure if format restrictions differ for src and dst, or if
 * they only matter when src fmt != dst fmt..  but there appear to
 * be *some* limitations so let's just start rejecting stuff that
 * piglit complains about
 */
static bool
ok_format(enum pipe_format fmt)
{
   if (util_format_is_compressed(fmt))
      return false;

   switch (fmt) {
   case PIPE_FORMAT_R10G10B10A2_SSCALED:
   case PIPE_FORMAT_R10G10B10A2_SNORM:
   case PIPE_FORMAT_B10G10R10A2_USCALED:
   case PIPE_FORMAT_B10G10R10A2_SSCALED:
   case PIPE_FORMAT_B10G10R10A2_SNORM:
   case PIPE_FORMAT_R10G10B10A2_UNORM:
   case PIPE_FORMAT_R10G10B10A2_USCALED:
   case PIPE_FORMAT_B10G10R10A2_UNORM:
   case PIPE_FORMAT_R10SG10SB10SA2U_NORM:
   case PIPE_FORMAT_B10G10R10A2_UINT:
   case PIPE_FORMAT_R10G10B10A2_UINT:
      return false;
   default:
      break;
   }

   if (fd5_pipe2color(fmt) == RB5_NONE)
      return false;

   return true;
}

static bool
can_do_blit(const struct pipe_blit_info *info)
{
   /* I think we can do scaling, but not in z dimension since that would
    * require blending..
    */
   if (info->dst.box.depth != info->src.box.depth)
      return false;

   if (!ok_format(info->dst.format))
      return false;

   if (!ok_format(info->src.format))
      return false;

   /* hw ignores {SRC,DST}_INFO.COLOR_SWAP if {SRC,DST}_INFO.TILE_MODE
    * is set (not linear).  We can kind of get around that when tiling/
    * untiling by setting both src and dst COLOR_SWAP=WZYX, but that
    * means the formats must match:
    */
   if ((fd_resource(info->dst.resource)->layout.tile_mode ||
        fd_resource(info->src.resource)->layout.tile_mode) &&
       info->dst.format != info->src.format)
      return false;

   /* until we figure out a few more registers: */
   if ((info->dst.box.width != info->src.box.width) ||
       (info->dst.box.height != info->src.box.height))
      return false;

   /* src box can be inverted, which we don't support.. dst box cannot: */
   if ((info->src.box.width < 0) || (info->src.box.height < 0))
      return false;

   if (!ok_dims(info->src.resource, &info->src.box, info->src.level))
      return false;

   if (!ok_dims(info->dst.resource, &info->dst.box, info->dst.level))
      return false;

   debug_assert(info->dst.box.width >= 0);
   debug_assert(info->dst.box.height >= 0);
   debug_assert(info->dst.box.depth >= 0);

   if ((info->dst.resource->nr_samples > 1) ||
       (info->src.resource->nr_samples > 1))
      return false;

   if (info->scissor_enable)
      return false;

   if (info->window_rectangle_include)
      return false;

   if (info->render_condition_enable)
      return false;

   if (info->alpha_blend)
      return false;

   if (info->filter != PIPE_TEX_FILTER_NEAREST)
      return false;

   if (info->mask != util_format_get_mask(info->src.format))
      return false;

   if (info->mask != util_format_get_mask(info->dst.format))
      return false;

   return true;
}

static void
emit_setup(struct fd_ringbuffer *ring)
{
   OUT_PKT4(ring, REG_A5XX_RB_RENDER_CNTL, 1);
   OUT_RING(ring, 0x00000008);

   OUT_PKT4(ring, REG_A5XX_UNKNOWN_2100, 1);
   OUT_RING(ring, 0x86000000); /* UNKNOWN_2100 */

   OUT_PKT4(ring, REG_A5XX_UNKNOWN_2180, 1);
   OUT_RING(ring, 0x86000000); /* UNKNOWN_2180 */

   OUT_PKT4(ring, REG_A5XX_UNKNOWN_2184, 1);
   OUT_RING(ring, 0x00000009); /* UNKNOWN_2184 */

   OUT_PKT4(ring, REG_A5XX_RB_CNTL, 1);
   OUT_RING(ring, A5XX_RB_CNTL_BYPASS);

   OUT_PKT4(ring, REG_A5XX_RB_MODE_CNTL, 1);
   OUT_RING(ring, 0x00000004); /* RB_MODE_CNTL */

   OUT_PKT4(ring, REG_A5XX_SP_MODE_CNTL, 1);
   OUT_RING(ring, 0x0000000c); /* SP_MODE_CNTL */

   OUT_PKT4(ring, REG_A5XX_TPL1_MODE_CNTL, 1);
   OUT_RING(ring, 0x00000344); /* TPL1_MODE_CNTL */

   OUT_PKT4(ring, REG_A5XX_HLSQ_MODE_CNTL, 1);
   OUT_RING(ring, 0x00000002); /* HLSQ_MODE_CNTL */

   OUT_PKT4(ring, REG_A5XX_GRAS_CL_CNTL, 1);
   OUT_RING(ring, 0x00000181); /* GRAS_CL_CNTL */
}

/* buffers need to be handled specially since x/width can exceed the bounds
 * supported by hw.. if necessary decompose into (potentially) two 2D blits
 */
static void
emit_blit_buffer(struct fd_ringbuffer *ring, const struct pipe_blit_info *info)
{
   const struct pipe_box *sbox = &info->src.box;
   const struct pipe_box *dbox = &info->dst.box;
   struct fd_resource *src, *dst;
   unsigned sshift, dshift;

   src = fd_resource(info->src.resource);
   dst = fd_resource(info->dst.resource);

   debug_assert(src->layout.cpp == 1);
   debug_assert(dst->layout.cpp == 1);
   debug_assert(info->src.resource->format == info->dst.resource->format);
   debug_assert((sbox->y == 0) && (sbox->height == 1));
   debug_assert((dbox->y == 0) && (dbox->height == 1));
   debug_assert((sbox->z == 0) && (sbox->depth == 1));
   debug_assert((dbox->z == 0) && (dbox->depth == 1));
   debug_assert(sbox->width == dbox->width);
   debug_assert(info->src.level == 0);
   debug_assert(info->dst.level == 0);

   /*
    * Buffers can have dimensions bigger than max width, remap into
    * multiple 1d blits to fit within max dimension
    *
    * Note that blob uses .ARRAY_PITCH=128 for blitting buffers, which
    * seems to prevent overfetch related faults.  Not quite sure what
    * the deal is there.
    *
    * Low 6 bits of SRC/DST addresses need to be zero (ie. address
    * aligned to 64) so we need to shift src/dst x1/x2 to make up the
    * difference.  On top of already splitting up the blit so width
    * isn't > 16k.
    *
    * We perhaps could do a bit better, if src and dst are aligned but
    * in the worst case this means we have to split the copy up into
    * 16k (0x4000) minus 64 (0x40).
    */

   sshift = sbox->x & 0x3f;
   dshift = dbox->x & 0x3f;

   for (unsigned off = 0; off < sbox->width; off += (0x4000 - 0x40)) {
      unsigned soff, doff, w, p;

      soff = (sbox->x + off) & ~0x3f;
      doff = (dbox->x + off) & ~0x3f;

      w = MIN2(sbox->width - off, (0x4000 - 0x40));
      p = align(w, 64);

      debug_assert((soff + w) <= fd_bo_size(src->bo));
      debug_assert((doff + w) <= fd_bo_size(dst->bo));

      OUT_PKT7(ring, CP_SET_RENDER_MODE, 1);
      OUT_RING(ring, CP_SET_RENDER_MODE_0_MODE(BLIT2D));

      /*
       * Emit source:
       */
      OUT_PKT4(ring, REG_A5XX_RB_2D_SRC_INFO, 9);
      OUT_RING(ring, A5XX_RB_2D_SRC_INFO_COLOR_FORMAT(RB5_R8_UNORM) |
                        A5XX_RB_2D_SRC_INFO_TILE_MODE(TILE5_LINEAR) |
                        A5XX_RB_2D_SRC_INFO_COLOR_SWAP(WZYX));
      OUT_RELOC(ring, src->bo, soff, 0, 0); /* RB_2D_SRC_LO/HI */
      OUT_RING(ring, A5XX_RB_2D_SRC_SIZE_PITCH(p) |
                        A5XX_RB_2D_SRC_SIZE_ARRAY_PITCH(128));
      OUT_RING(ring, 0x00000000);
      OUT_RING(ring, 0x00000000);
      OUT_RING(ring, 0x00000000);
      OUT_RING(ring, 0x00000000);
      OUT_RING(ring, 0x00000000);

      OUT_PKT4(ring, REG_A5XX_GRAS_2D_SRC_INFO, 1);
      OUT_RING(ring, A5XX_GRAS_2D_SRC_INFO_COLOR_FORMAT(RB5_R8_UNORM) |
                        A5XX_GRAS_2D_SRC_INFO_COLOR_SWAP(WZYX));

      /*
       * Emit destination:
       */
      OUT_PKT4(ring, REG_A5XX_RB_2D_DST_INFO, 9);
      OUT_RING(ring, A5XX_RB_2D_DST_INFO_COLOR_FORMAT(RB5_R8_UNORM) |
                        A5XX_RB_2D_DST_INFO_TILE_MODE(TILE5_LINEAR) |
                        A5XX_RB_2D_DST_INFO_COLOR_SWAP(WZYX));
      OUT_RELOC(ring, dst->bo, doff, 0, 0); /* RB_2D_DST_LO/HI */
      OUT_RING(ring, A5XX_RB_2D_DST_SIZE_PITCH(p) |
                        A5XX_RB_2D_DST_SIZE_ARRAY_PITCH(128));
      OUT_RING(ring, 0x00000000);
      OUT_RING(ring, 0x00000000);
      OUT_RING(ring, 0x00000000);
      OUT_RING(ring, 0x00000000);
      OUT_RING(ring, 0x00000000);

      OUT_PKT4(ring, REG_A5XX_GRAS_2D_DST_INFO, 1);
      OUT_RING(ring, A5XX_GRAS_2D_DST_INFO_COLOR_FORMAT(RB5_R8_UNORM) |
                        A5XX_GRAS_2D_DST_INFO_COLOR_SWAP(WZYX));

      /*
       * Blit command:
       */
      OUT_PKT7(ring, CP_BLIT, 5);
      OUT_RING(ring, CP_BLIT_0_OP(BLIT_OP_COPY));
      OUT_RING(ring, CP_BLIT_1_SRC_X1(sshift) | CP_BLIT_1_SRC_Y1(0));
      OUT_RING(ring, CP_BLIT_2_SRC_X2(sshift + w - 1) | CP_BLIT_2_SRC_Y2(0));
      OUT_RING(ring, CP_BLIT_3_DST_X1(dshift) | CP_BLIT_3_DST_Y1(0));
      OUT_RING(ring, CP_BLIT_4_DST_X2(dshift + w - 1) | CP_BLIT_4_DST_Y2(0));

      OUT_PKT7(ring, CP_SET_RENDER_MODE, 1);
      OUT_RING(ring, CP_SET_RENDER_MODE_0_MODE(END2D));

      OUT_WFI5(ring);
   }
}

static void
emit_blit(struct fd_ringbuffer *ring, const struct pipe_blit_info *info)
{
   const struct pipe_box *sbox = &info->src.box;
   const struct pipe_box *dbox = &info->dst.box;
   struct fd_resource *src, *dst;
   struct fdl_slice *sslice, *dslice;
   enum a5xx_color_fmt sfmt, dfmt;
   enum a5xx_tile_mode stile, dtile;
   enum a3xx_color_swap sswap, dswap;
   unsigned ssize, dsize, spitch, dpitch;
   unsigned sx1, sy1, sx2, sy2;
   unsigned dx1, dy1, dx2, dy2;

   src = fd_resource(info->src.resource);
   dst = fd_resource(info->dst.resource);

   sslice = fd_resource_slice(src, info->src.level);
   dslice = fd_resource_slice(dst, info->dst.level);

   sfmt = fd5_pipe2color(info->src.format);
   dfmt = fd5_pipe2color(info->dst.format);

   stile = fd_resource_tile_mode(info->src.resource, info->src.level);
   dtile = fd_resource_tile_mode(info->dst.resource, info->dst.level);

   sswap = fd5_pipe2swap(info->src.format);
   dswap = fd5_pipe2swap(info->dst.format);

   spitch = fd_resource_pitch(src, info->src.level);
   dpitch = fd_resource_pitch(dst, info->dst.level);

   /* if dtile, then dswap ignored by hw, and likewise if stile then sswap
    * ignored by hw.. but in this case we have already rejected the blit
    * if src and dst formats differ, so juse use WZYX for both src and
    * dst swap mode (so we don't change component order)
    */
   if (stile || dtile) {
      debug_assert(info->src.format == info->dst.format);
      sswap = dswap = WZYX;
   }

   sx1 = sbox->x;
   sy1 = sbox->y;
   sx2 = sbox->x + sbox->width - 1;
   sy2 = sbox->y + sbox->height - 1;

   dx1 = dbox->x;
   dy1 = dbox->y;
   dx2 = dbox->x + dbox->width - 1;
   dy2 = dbox->y + dbox->height - 1;

   if (info->src.resource->target == PIPE_TEXTURE_3D)
      ssize = sslice->size0;
   else
      ssize = src->layout.layer_size;

   if (info->dst.resource->target == PIPE_TEXTURE_3D)
      dsize = dslice->size0;
   else
      dsize = dst->layout.layer_size;

   for (unsigned i = 0; i < info->dst.box.depth; i++) {
      unsigned soff = fd_resource_offset(src, info->src.level, sbox->z + i);
      unsigned doff = fd_resource_offset(dst, info->dst.level, dbox->z + i);

      debug_assert((soff + (sbox->height * spitch)) <= fd_bo_size(src->bo));
      debug_assert((doff + (dbox->height * dpitch)) <= fd_bo_size(dst->bo));

      OUT_PKT7(ring, CP_SET_RENDER_MODE, 1);
      OUT_RING(ring, CP_SET_RENDER_MODE_0_MODE(BLIT2D));

      /*
       * Emit source:
       */
      OUT_PKT4(ring, REG_A5XX_RB_2D_SRC_INFO, 9);
      OUT_RING(ring, A5XX_RB_2D_SRC_INFO_COLOR_FORMAT(sfmt) |
                        A5XX_RB_2D_SRC_INFO_TILE_MODE(stile) |
                        A5XX_RB_2D_SRC_INFO_COLOR_SWAP(sswap));
      OUT_RELOC(ring, src->bo, soff, 0, 0); /* RB_2D_SRC_LO/HI */
      OUT_RING(ring, A5XX_RB_2D_SRC_SIZE_PITCH(spitch) |
                        A5XX_RB_2D_SRC_SIZE_ARRAY_PITCH(ssize));
      OUT_RING(ring, 0x00000000);
      OUT_RING(ring, 0x00000000);
      OUT_RING(ring, 0x00000000);
      OUT_RING(ring, 0x00000000);
      OUT_RING(ring, 0x00000000);

      OUT_PKT4(ring, REG_A5XX_GRAS_2D_SRC_INFO, 1);
      OUT_RING(ring, A5XX_GRAS_2D_SRC_INFO_COLOR_FORMAT(sfmt) |
                        A5XX_GRAS_2D_SRC_INFO_TILE_MODE(stile) |
                        A5XX_GRAS_2D_SRC_INFO_COLOR_SWAP(sswap));

      /*
       * Emit destination:
       */
      OUT_PKT4(ring, REG_A5XX_RB_2D_DST_INFO, 9);
      OUT_RING(ring, A5XX_RB_2D_DST_INFO_COLOR_FORMAT(dfmt) |
                        A5XX_RB_2D_DST_INFO_TILE_MODE(dtile) |
                        A5XX_RB_2D_DST_INFO_COLOR_SWAP(dswap));
      OUT_RELOC(ring, dst->bo, doff, 0, 0); /* RB_2D_DST_LO/HI */
      OUT_RING(ring, A5XX_RB_2D_DST_SIZE_PITCH(dpitch) |
                        A5XX_RB_2D_DST_SIZE_ARRAY_PITCH(dsize));
      OUT_RING(ring, 0x00000000);
      OUT_RING(ring, 0x00000000);
      OUT_RING(ring, 0x00000000);
      OUT_RING(ring, 0x00000000);
      OUT_RING(ring, 0x00000000);

      OUT_PKT4(ring, REG_A5XX_GRAS_2D_DST_INFO, 1);
      OUT_RING(ring, A5XX_GRAS_2D_DST_INFO_COLOR_FORMAT(dfmt) |
                        A5XX_GRAS_2D_DST_INFO_TILE_MODE(dtile) |
                        A5XX_GRAS_2D_DST_INFO_COLOR_SWAP(dswap));

      /*
       * Blit command:
       */
      OUT_PKT7(ring, CP_BLIT, 5);
      OUT_RING(ring, CP_BLIT_0_OP(BLIT_OP_COPY));
      OUT_RING(ring, CP_BLIT_1_SRC_X1(sx1) | CP_BLIT_1_SRC_Y1(sy1));
      OUT_RING(ring, CP_BLIT_2_SRC_X2(sx2) | CP_BLIT_2_SRC_Y2(sy2));
      OUT_RING(ring, CP_BLIT_3_DST_X1(dx1) | CP_BLIT_3_DST_Y1(dy1));
      OUT_RING(ring, CP_BLIT_4_DST_X2(dx2) | CP_BLIT_4_DST_Y2(dy2));

      OUT_PKT7(ring, CP_SET_RENDER_MODE, 1);
      OUT_RING(ring, CP_SET_RENDER_MODE_0_MODE(END2D));
   }
}

bool
fd5_blitter_blit(struct fd_context *ctx,
                 const struct pipe_blit_info *info) assert_dt
{
   struct fd_batch *batch;

   if (!can_do_blit(info)) {
      return false;
   }

   struct fd_resource *src = fd_resource(info->src.resource);
   struct fd_resource *dst = fd_resource(info->dst.resource);

   batch = fd_bc_alloc_batch(ctx, true);

   fd_screen_lock(ctx->screen);

   fd_batch_resource_read(batch, src);
   fd_batch_resource_write(batch, dst);

   fd_screen_unlock(ctx->screen);

   DBG_BLIT(info, batch);

   fd_batch_update_queries(batch);

   emit_setup(batch->draw);

   if ((info->src.resource->target == PIPE_BUFFER) &&
       (info->dst.resource->target == PIPE_BUFFER)) {
      assert(fd_resource(info->src.resource)->layout.tile_mode == TILE5_LINEAR);
      assert(fd_resource(info->dst.resource)->layout.tile_mode == TILE5_LINEAR);
      emit_blit_buffer(batch->draw, info);
   } else {
      /* I don't *think* we need to handle blits between buffer <-> !buffer */
      debug_assert(info->src.resource->target != PIPE_BUFFER);
      debug_assert(info->dst.resource->target != PIPE_BUFFER);
      emit_blit(batch->draw, info);
   }

   fd_batch_needs_flush(batch);

   fd_batch_flush(batch);
   fd_batch_reference(&batch, NULL);

   /* Acc query state will have been dirtied by our fd_batch_update_queries, so
    * the ctx->batch may need to turn its queries back on.
    */
   ctx->update_active_queries = true;

   return true;
}

unsigned
fd5_tile_mode(const struct pipe_resource *tmpl)
{
   /* basically just has to be a format we can blit, so uploads/downloads
    * via linear staging buffer works:
    */
   if (ok_format(tmpl->format))
      return TILE5_3;

   return TILE5_LINEAR;
}