freedreno: proper locking for iterating dependent batches
[mesa.git] / src / gallium / drivers / radeonsi / cik_sdma.c
index fdcf22f6d4968c54a6698d02c230e8d8f5642425..3e3b9cd4329f344b5f2bac472eee4b484d5552ca 100644 (file)
@@ -20,9 +20,6 @@
  * 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:
- *      Jerome Glisse
  */
 
 #include "sid.h"
@@ -50,14 +47,14 @@ static void cik_sdma_copy_buffer(struct si_context *ctx,
        src_offset += rsrc->gpu_address;
 
        ncopy = DIV_ROUND_UP(size, CIK_SDMA_COPY_MAX_SIZE);
-       r600_need_dma_space(&ctx->b, ncopy * 7, rdst, rsrc);
+       si_need_dma_space(&ctx->b, ncopy * 7, rdst, rsrc);
 
        for (i = 0; i < ncopy; i++) {
                csize = MIN2(size, CIK_SDMA_COPY_MAX_SIZE);
                radeon_emit(cs, CIK_SDMA_PACKET(CIK_SDMA_OPCODE_COPY,
                                                CIK_SDMA_COPY_SUB_OPCODE_LINEAR,
                                                0));
-               radeon_emit(cs, csize);
+               radeon_emit(cs, ctx->b.chip_class >= GFX9 ? csize - 1 : csize);
                radeon_emit(cs, 0); /* src/dst endian swap */
                radeon_emit(cs, src_offset);
                radeon_emit(cs, src_offset >> 32);
@@ -80,7 +77,8 @@ static void cik_sdma_clear_buffer(struct pipe_context *ctx,
        unsigned i, ncopy, csize;
        struct r600_resource *rdst = r600_resource(dst);
 
-       if (!cs || offset % 4 != 0 || size % 4 != 0) {
+       if (!cs || offset % 4 != 0 || size % 4 != 0 ||
+           dst->flags & PIPE_RESOURCE_FLAG_SPARSE) {
                ctx->clear_buffer(ctx, dst, offset, size, &clear_value, 4);
                return;
        }
@@ -94,7 +92,7 @@ static void cik_sdma_clear_buffer(struct pipe_context *ctx,
 
        /* the same maximum size as for copying */
        ncopy = DIV_ROUND_UP(size, CIK_SDMA_COPY_MAX_SIZE);
-       r600_need_dma_space(&sctx->b, ncopy * 5, rdst, NULL);
+       si_need_dma_space(&sctx->b, ncopy * 5, rdst, NULL);
 
        for (i = 0; i < ncopy; i++) {
                csize = MIN2(size, CIK_SDMA_COPY_MAX_SIZE);
@@ -103,7 +101,7 @@ static void cik_sdma_clear_buffer(struct pipe_context *ctx,
                radeon_emit(cs, offset);
                radeon_emit(cs, offset >> 32);
                radeon_emit(cs, clear_value);
-               radeon_emit(cs, csize);
+               radeon_emit(cs, sctx->b.chip_class >= GFX9 ? csize - 1 : csize);
                offset += csize;
                size -= csize;
        }
@@ -119,9 +117,9 @@ static unsigned encode_tile_info(struct si_context *sctx,
                                 struct r600_texture *tex, unsigned level,
                                 bool set_bpp)
 {
-       struct radeon_info *info = &sctx->screen->b.info;
-       unsigned tile_index = tex->surface.tiling_index[level];
-       unsigned macro_tile_index = tex->surface.macro_tile_index;
+       struct radeon_info *info = &sctx->screen->info;
+       unsigned tile_index = tex->surface.u.legacy.tiling_index[level];
+       unsigned macro_tile_index = tex->surface.u.legacy.macro_tile_index;
        unsigned tile_mode = info->si_tile_mode_array[tile_index];
        unsigned macro_tile_mode = info->cik_macrotile_mode_array[macro_tile_index];
 
@@ -129,7 +127,7 @@ static unsigned encode_tile_info(struct si_context *sctx,
                (G_009910_ARRAY_MODE(tile_mode) << 3) |
                (G_009910_MICRO_TILE_MODE_NEW(tile_mode) << 8) |
                /* Non-depth modes don't have TILE_SPLIT set. */
-               ((util_logbase2(tex->surface.tile_split >> 6)) << 11) |
+               ((util_logbase2(tex->surface.u.legacy.tile_split >> 6)) << 11) |
                (G_009990_BANK_WIDTH(macro_tile_mode) << 15) |
                (G_009990_BANK_HEIGHT(macro_tile_mode) << 18) |
                (G_009990_NUM_BANKS(macro_tile_mode) << 21) |
@@ -145,26 +143,30 @@ static bool cik_sdma_copy_texture(struct si_context *sctx,
                                  unsigned src_level,
                                  const struct pipe_box *src_box)
 {
-       struct radeon_info *info = &sctx->screen->b.info;
+       struct radeon_info *info = &sctx->screen->info;
        struct r600_texture *rsrc = (struct r600_texture*)src;
        struct r600_texture *rdst = (struct r600_texture*)dst;
        unsigned bpp = rdst->surface.bpe;
        uint64_t dst_address = rdst->resource.gpu_address +
-                              rdst->surface.level[dst_level].offset;
+                              rdst->surface.u.legacy.level[dst_level].offset;
        uint64_t src_address = rsrc->resource.gpu_address +
-                              rsrc->surface.level[src_level].offset;
-       unsigned dst_mode = rdst->surface.level[dst_level].mode;
-       unsigned src_mode = rsrc->surface.level[src_level].mode;
-       unsigned dst_tile_index = rdst->surface.tiling_index[dst_level];
-       unsigned src_tile_index = rsrc->surface.tiling_index[src_level];
+                              rsrc->surface.u.legacy.level[src_level].offset;
+       unsigned dst_mode = rdst->surface.u.legacy.level[dst_level].mode;
+       unsigned src_mode = rsrc->surface.u.legacy.level[src_level].mode;
+       unsigned dst_tile_index = rdst->surface.u.legacy.tiling_index[dst_level];
+       unsigned src_tile_index = rsrc->surface.u.legacy.tiling_index[src_level];
        unsigned dst_tile_mode = info->si_tile_mode_array[dst_tile_index];
        unsigned src_tile_mode = info->si_tile_mode_array[src_tile_index];
        unsigned dst_micro_mode = G_009910_MICRO_TILE_MODE_NEW(dst_tile_mode);
        unsigned src_micro_mode = G_009910_MICRO_TILE_MODE_NEW(src_tile_mode);
-       unsigned dst_pitch = rdst->surface.level[dst_level].nblk_x;
-       unsigned src_pitch = rsrc->surface.level[src_level].nblk_x;
-       uint64_t dst_slice_pitch = rdst->surface.level[dst_level].slice_size / bpp;
-       uint64_t src_slice_pitch = rsrc->surface.level[src_level].slice_size / bpp;
+       unsigned dst_tile_swizzle = dst_mode == RADEON_SURF_MODE_2D ?
+                                           rdst->surface.tile_swizzle : 0;
+       unsigned src_tile_swizzle = src_mode == RADEON_SURF_MODE_2D ?
+                                           rsrc->surface.tile_swizzle : 0;
+       unsigned dst_pitch = rdst->surface.u.legacy.level[dst_level].nblk_x;
+       unsigned src_pitch = rsrc->surface.u.legacy.level[src_level].nblk_x;
+       uint64_t dst_slice_pitch = ((uint64_t)rdst->surface.u.legacy.level[dst_level].slice_size_dw * 4) / bpp;
+       uint64_t src_slice_pitch = ((uint64_t)rsrc->surface.u.legacy.level[src_level].slice_size_dw * 4) / bpp;
        unsigned dst_width = minify_as_blocks(rdst->resource.b.b.width0,
                                              dst_level, rdst->surface.blk_w);
        unsigned src_width = minify_as_blocks(rsrc->resource.b.b.width0,
@@ -182,14 +184,14 @@ static bool cik_sdma_copy_texture(struct si_context *sctx,
 
        assert(src_level <= src->last_level);
        assert(dst_level <= dst->last_level);
-       assert(rdst->surface.level[dst_level].offset +
+       assert(rdst->surface.u.legacy.level[dst_level].offset +
               dst_slice_pitch * bpp * (dstz + src_box->depth) <=
               rdst->resource.buf->size);
-       assert(rsrc->surface.level[src_level].offset +
+       assert(rsrc->surface.u.legacy.level[src_level].offset +
               src_slice_pitch * bpp * (srcz + src_box->depth) <=
               rsrc->resource.buf->size);
 
-       if (!r600_prepare_for_dma_blit(&sctx->b, rdst, dst_level, dstx, dsty,
+       if (!si_prepare_for_dma_blit(&sctx->b, rdst, dst_level, dstx, dsty,
                                        dstz, rsrc, src_level, src_box))
                return false;
 
@@ -204,6 +206,9 @@ static bool cik_sdma_copy_texture(struct si_context *sctx,
            dstz >= (1 << 11))
                return false;
 
+       dst_address |= dst_tile_swizzle << 8;
+       src_address |= src_tile_swizzle << 8;
+
        /* Linear -> linear sub-window copy. */
        if (dst_mode == RADEON_SURF_MODE_LINEAR_ALIGNED &&
            src_mode == RADEON_SURF_MODE_LINEAR_ALIGNED &&
@@ -227,7 +232,7 @@ static bool cik_sdma_copy_texture(struct si_context *sctx,
              srcy + copy_height != (1 << 14)))) {
                struct radeon_winsys_cs *cs = sctx->b.dma.cs;
 
-               r600_need_dma_space(&sctx->b, 13, &rdst->resource, &rsrc->resource);
+               si_need_dma_space(&sctx->b, 13, &rdst->resource, &rsrc->resource);
 
                radeon_emit(cs, CIK_SDMA_PACKET(CIK_SDMA_OPCODE_COPY,
                                                CIK_SDMA_COPY_SUB_OPCODE_LINEAR_SUB_WINDOW, 0) |
@@ -350,14 +355,14 @@ static bool cik_sdma_copy_texture(struct si_context *sctx,
                 * starts reading from an address preceding linear_address!!!
                 */
                start_linear_address =
-                       linear->surface.level[linear_level].offset +
+                       linear->surface.u.legacy.level[linear_level].offset +
                        bpp * (linear_z * linear_slice_pitch +
                               linear_y * linear_pitch +
                               linear_x);
                start_linear_address -= (int)(bpp * (tiled_x % granularity));
 
                end_linear_address =
-                       linear->surface.level[linear_level].offset +
+                       linear->surface.u.legacy.level[linear_level].offset +
                        bpp * ((linear_z + copy_depth - 1) * linear_slice_pitch +
                               (linear_y + copy_height - 1) * linear_pitch +
                               (linear_x + copy_width));
@@ -379,7 +384,7 @@ static bool cik_sdma_copy_texture(struct si_context *sctx,
                    copy_width_aligned % xalign == 0 &&
                    tiled_micro_mode != V_009910_ADDR_SURF_ROTATED_MICRO_TILING &&
                    /* check if everything fits into the bitfields */
-                   tiled->surface.tile_split <= 4096 &&
+                   tiled->surface.u.legacy.tile_split <= 4096 &&
                    pitch_tile_max < (1 << 11) &&
                    slice_tile_max < (1 << 22) &&
                    linear_pitch <= (1 << 14) &&
@@ -390,7 +395,7 @@ static bool cik_sdma_copy_texture(struct si_context *sctx,
                        struct radeon_winsys_cs *cs = sctx->b.dma.cs;
                        uint32_t direction = linear == rdst ? 1u << 31 : 0;
 
-                       r600_need_dma_space(&sctx->b, 14, &rdst->resource, &rsrc->resource);
+                       si_need_dma_space(&sctx->b, 14, &rdst->resource, &rsrc->resource);
 
                        radeon_emit(cs, CIK_SDMA_PACKET(CIK_SDMA_OPCODE_COPY,
                                                        CIK_SDMA_COPY_SUB_OPCODE_TILED_SUB_WINDOW, 0) |
@@ -423,8 +428,8 @@ static bool cik_sdma_copy_texture(struct si_context *sctx,
            /* check if these fit into the bitfields */
            src_address % 256 == 0 &&
            dst_address % 256 == 0 &&
-           rsrc->surface.tile_split <= 4096 &&
-           rdst->surface.tile_split <= 4096 &&
+           rsrc->surface.u.legacy.tile_split <= 4096 &&
+           rdst->surface.u.legacy.tile_split <= 4096 &&
            dstx % 8 == 0 &&
            dsty % 8 == 0 &&
            srcx % 8 == 0 &&
@@ -484,7 +489,7 @@ static bool cik_sdma_copy_texture(struct si_context *sctx,
                      dstx + copy_width != (1 << 14)))) {
                        struct radeon_winsys_cs *cs = sctx->b.dma.cs;
 
-                       r600_need_dma_space(&sctx->b, 15, &rdst->resource, &rsrc->resource);
+                       si_need_dma_space(&sctx->b, 15, &rdst->resource, &rsrc->resource);
 
                        radeon_emit(cs, CIK_SDMA_PACKET(CIK_SDMA_OPCODE_COPY,
                                                        CIK_SDMA_COPY_SUB_OPCODE_T2T_SUB_WINDOW, 0));
@@ -526,7 +531,9 @@ static void cik_sdma_copy(struct pipe_context *ctx,
 {
        struct si_context *sctx = (struct si_context *)ctx;
 
-       if (!sctx->b.dma.cs)
+       if (!sctx->b.dma.cs ||
+           src->flags & PIPE_RESOURCE_FLAG_SPARSE ||
+           dst->flags & PIPE_RESOURCE_FLAG_SPARSE)
                goto fallback;
 
        if (dst->target == PIPE_BUFFER && src->target == PIPE_BUFFER) {
@@ -534,7 +541,8 @@ static void cik_sdma_copy(struct pipe_context *ctx,
                return;
        }
 
-       if (cik_sdma_copy_texture(sctx, dst, dst_level, dstx, dsty, dstz,
+       if ((sctx->b.chip_class == CIK || sctx->b.chip_class == VI) &&
+           cik_sdma_copy_texture(sctx, dst, dst_level, dstx, dsty, dstz,
                                  src, src_level, src_box))
                return;