freedreno: add resource tracking support for written buffers
[mesa.git] / src / gallium / drivers / freedreno / freedreno_draw.c
index d1f1c43d7b4f58e01d064db8a72d98b8ec8bc984..ae75b3efdccf1db1ce8733f2ffa2f97ad2ea1742 100644 (file)
 #include "freedreno_context.h"
 #include "freedreno_state.h"
 #include "freedreno_resource.h"
+#include "freedreno_query_hw.h"
 #include "freedreno_util.h"
 
-
-static enum pc_di_primtype
-mode2primtype(unsigned mode)
-{
-       switch (mode) {
-       case PIPE_PRIM_POINTS:         return DI_PT_POINTLIST;
-       case PIPE_PRIM_LINES:          return DI_PT_LINELIST;
-       case PIPE_PRIM_LINE_STRIP:     return DI_PT_LINESTRIP;
-       case PIPE_PRIM_TRIANGLES:      return DI_PT_TRILIST;
-       case PIPE_PRIM_TRIANGLE_STRIP: return DI_PT_TRISTRIP;
-       case PIPE_PRIM_TRIANGLE_FAN:   return DI_PT_TRIFAN;
-       case PIPE_PRIM_QUADS:          return DI_PT_QUADLIST;
-       case PIPE_PRIM_QUAD_STRIP:     return DI_PT_QUADSTRIP;
-       case PIPE_PRIM_POLYGON:        return DI_PT_POLYGON;
-       }
-       DBG("unsupported mode: (%s) %d", u_prim_name(mode), mode);
-       assert(0);
-       return DI_PT_NONE;
-}
-
-static enum pc_di_index_size
-size2indextype(unsigned index_size)
+static void
+resource_used(struct fd_context *ctx, struct pipe_resource *prsc, boolean reading)
 {
-       switch (index_size) {
-       case 1: return INDEX_SIZE_8_BIT;
-       case 2: return INDEX_SIZE_16_BIT;
-       case 4: return INDEX_SIZE_32_BIT;
-       }
-       DBG("unsupported index size: %d", index_size);
-       assert(0);
-       return INDEX_SIZE_IGN;
-}
+       struct fd_resource *rsc;
 
-/* this is same for a2xx/a3xx, so split into helper: */
-void
-fd_draw_emit(struct fd_context *ctx, const struct pipe_draw_info *info)
-{
-       struct pipe_index_buffer *idx = &ctx->indexbuf;
-       struct fd_bo *idx_bo = NULL;
-       enum pc_di_index_size idx_type = INDEX_SIZE_IGN;
-       enum pc_di_src_sel src_sel;
-       uint32_t idx_size, idx_offset;
-
-       if (info->indexed) {
-               assert(!idx->user_buffer);
-
-               idx_bo = fd_resource(idx->buffer)->bo;
-               idx_type = size2indextype(idx->index_size);
-               idx_size = idx->index_size * info->count;
-               idx_offset = idx->offset;
-               src_sel = DI_SRC_SEL_DMA;
-       } else {
-               idx_bo = NULL;
-               idx_type = INDEX_SIZE_IGN;
-               idx_size = 0;
-               idx_offset = 0;
-               src_sel = DI_SRC_SEL_AUTO_INDEX;
-       }
+       if (!prsc)
+               return;
 
-       fd_draw(ctx, mode2primtype(info->mode), src_sel, info->count,
-                       idx_type, idx_size, idx_offset, idx_bo);
+       rsc = fd_resource(prsc);
+       if (reading)
+               rsc->reading = true;
+       else
+               rsc->writing = true;
+       list_delinit(&rsc->list);
+       list_addtail(&rsc->list, &ctx->used_resources);
 }
 
 static void
@@ -115,6 +70,14 @@ fd_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
                return;
        }
 
+       /* emulate unsupported primitives: */
+       if (!fd_supported_prim(ctx, info->mode)) {
+               util_primconvert_save_index_buffer(ctx->primconvert, &ctx->indexbuf);
+               util_primconvert_save_rasterizer_state(ctx->primconvert, ctx->rasterizer);
+               util_primconvert_draw_vbo(ctx->primconvert, info);
+               return;
+       }
+
        ctx->needs_flush = true;
 
        /*
@@ -128,8 +91,12 @@ fd_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
        }
 
        if (fd_stencil_enabled(ctx)) {
+               struct fd_resource *rsc = fd_resource(pfb->zsbuf->texture);
                buffers |= FD_BUFFER_STENCIL;
-               fd_resource(pfb->zsbuf->texture)->dirty = true;
+               if (rsc->stencil)
+                       rsc->stencil->dirty = true;
+               else
+                       rsc->dirty = true;
                ctx->gmem_reason |= FD_GMEM_STENCIL_ENABLED;
        }
 
@@ -137,10 +104,15 @@ fd_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
                ctx->gmem_reason |= FD_GMEM_LOGICOP_ENABLED;
 
        for (i = 0; i < pfb->nr_cbufs; i++) {
-               struct pipe_resource *surf = pfb->cbufs[i]->texture;
+               struct pipe_resource *surf;
+
+               if (!pfb->cbufs[i])
+                       continue;
+
+               surf = pfb->cbufs[i]->texture;
 
                fd_resource(surf)->dirty = true;
-               buffers |= FD_BUFFER_COLOR;
+               buffers |= PIPE_CLEAR_COLOR0 << i;
 
                if (surf->nr_samples > 1)
                        ctx->gmem_reason |= FD_GMEM_MSAA_ENABLED;
@@ -149,14 +121,59 @@ fd_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
                        ctx->gmem_reason |= FD_GMEM_BLEND_ENABLED;
        }
 
+       /* Skip over buffer 0, that is sent along with the command stream */
+       for (i = 1; i < PIPE_MAX_CONSTANT_BUFFERS; i++) {
+               resource_used(ctx, ctx->constbuf[PIPE_SHADER_VERTEX].cb[i].buffer, true);
+               resource_used(ctx, ctx->constbuf[PIPE_SHADER_FRAGMENT].cb[i].buffer, true);
+       }
+
+       /* Mark VBOs as being read */
+       for (i = 0; i < ctx->vtx.vertexbuf.count; i++) {
+               assert(!ctx->vtx.vertexbuf.vb[i].user_buffer);
+               resource_used(ctx, ctx->vtx.vertexbuf.vb[i].buffer, true);
+       }
+
+       /* Mark index buffer as being read */
+       resource_used(ctx, ctx->indexbuf.buffer, true);
+
+       /* Mark textures as being read */
+       for (i = 0; i < ctx->verttex.num_textures; i++)
+               if (ctx->verttex.textures[i])
+                       resource_used(ctx, ctx->verttex.textures[i]->texture, true);
+       for (i = 0; i < ctx->fragtex.num_textures; i++)
+               if (ctx->fragtex.textures[i])
+                       resource_used(ctx, ctx->fragtex.textures[i]->texture, true);
+
        ctx->num_draws++;
 
-       /* any buffers that haven't been cleared, we need to restore: */
+       ctx->stats.draw_calls++;
+       ctx->stats.prims_emitted +=
+               u_reduced_prims_for_vertices(info->mode, info->count);
+
+       /* any buffers that haven't been cleared yet, we need to restore: */
        ctx->restore |= buffers & (FD_BUFFER_ALL & ~ctx->cleared);
        /* and any buffers used, need to be resolved: */
        ctx->resolve |= buffers;
 
-       ctx->draw(ctx, info);
+       DBG("%x num_draws=%u (%s/%s)", buffers, ctx->num_draws,
+               util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
+               util_format_short_name(pipe_surface_format(pfb->zsbuf)));
+
+       fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_DRAW);
+       ctx->draw_vbo(ctx, info);
+
+       /* if an app (or, well, piglit test) does many thousands of draws
+        * without flush (or anything which implicitly flushes, like
+        * changing render targets), we can exceed the ringbuffer size.
+        * Since we don't currently have a sane way to wrapparound, and
+        * we use the same buffer for both draw and tiling commands, for
+        * now we need to do this hack and trigger flush if we are running
+        * low on remaining space for cmds:
+        */
+       if (((ctx->ring->cur - ctx->ring->start) >
+                               (ctx->ring->size/4 - FD_TILING_COMMANDS_DWORDS)) ||
+                       (fd_mesa_debug & FD_DBG_FLUSH))
+               fd_context_render(pctx);
 }
 
 /* TODO figure out how to make better use of existing state mechanism
@@ -171,16 +188,46 @@ fd_clear(struct pipe_context *pctx, unsigned buffers,
 {
        struct fd_context *ctx = fd_context(pctx);
        struct pipe_framebuffer_state *pfb = &ctx->framebuffer;
+       struct pipe_scissor_state *scissor = fd_context_get_scissor(ctx);
+       unsigned cleared_buffers;
+       int i;
+
+       /* for bookkeeping about which buffers have been cleared (and thus
+        * can fully or partially skip mem2gmem) we need to ignore buffers
+        * that have already had a draw, in case apps do silly things like
+        * clear after draw (ie. if you only clear the color buffer, but
+        * something like alpha-test causes side effects from the draw in
+        * the depth buffer, etc)
+        */
+       cleared_buffers = buffers & (FD_BUFFER_ALL & ~ctx->restore);
 
-       ctx->cleared |= buffers;
+       /* do we have full-screen scissor? */
+       if (!memcmp(scissor, &ctx->disabled_scissor, sizeof(*scissor))) {
+               ctx->cleared |= cleared_buffers;
+       } else {
+               ctx->partial_cleared |= cleared_buffers;
+               if (cleared_buffers & PIPE_CLEAR_COLOR)
+                       ctx->cleared_scissor.color = *scissor;
+               if (cleared_buffers & PIPE_CLEAR_DEPTH)
+                       ctx->cleared_scissor.depth = *scissor;
+               if (cleared_buffers & PIPE_CLEAR_STENCIL)
+                       ctx->cleared_scissor.stencil = *scissor;
+       }
        ctx->resolve |= buffers;
        ctx->needs_flush = true;
 
        if (buffers & PIPE_CLEAR_COLOR)
-               fd_resource(pfb->cbufs[0]->texture)->dirty = true;
+               for (i = 0; i < pfb->nr_cbufs; i++)
+                       if (buffers & (PIPE_CLEAR_COLOR0 << i))
+                               fd_resource(pfb->cbufs[i]->texture)->dirty = true;
 
        if (buffers & (PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL)) {
-               fd_resource(pfb->zsbuf->texture)->dirty = true;
+               struct fd_resource *rsc = fd_resource(pfb->zsbuf->texture);
+               if (rsc->stencil && buffers & PIPE_CLEAR_STENCIL)
+                       rsc->stencil->dirty = true;
+               if (!rsc->stencil || buffers & PIPE_CLEAR_DEPTH)
+                       rsc->dirty = true;
+
                ctx->gmem_reason |= FD_GMEM_CLEARS_DEPTH_STENCIL;
        }
 
@@ -188,9 +235,12 @@ fd_clear(struct pipe_context *pctx, unsigned buffers,
                util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
                util_format_short_name(pipe_surface_format(pfb->zsbuf)));
 
+       fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_CLEAR);
+
        ctx->clear(ctx, buffers, color, depth, stencil);
 
        ctx->dirty |= FD_DIRTY_ZSA |
+                       FD_DIRTY_VIEWPORT |
                        FD_DIRTY_RASTERIZER |
                        FD_DIRTY_SAMPLE_MASK |
                        FD_DIRTY_PROG |
@@ -221,6 +271,8 @@ fd_clear_depth_stencil(struct pipe_context *pctx, struct pipe_surface *ps,
 void
 fd_draw_init(struct pipe_context *pctx)
 {
+       list_inithead(&fd_context(pctx)->used_resources);
+
        pctx->draw_vbo = fd_draw_vbo;
        pctx->clear = fd_clear;
        pctx->clear_render_target = fd_clear_render_target;