freedreno/a6xx: cleanup magic registers
[mesa.git] / src / gallium / drivers / freedreno / freedreno_draw.c
index a4a19750a420dafa27a4313d07c2dc67f248e526..059de2ec8c612dcf9f1165bd4adc834f3b5a73a7 100644 (file)
@@ -1,5 +1,3 @@
-/* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
-
 /*
  * Copyright (C) 2012 Rob Clark <robclark@freedesktop.org>
  *
  */
 
 #include "pipe/p_state.h"
+#include "util/u_draw.h"
 #include "util/u_string.h"
 #include "util/u_memory.h"
 #include "util/u_prim.h"
 #include "util/u_format.h"
 #include "util/u_helpers.h"
 
+#include "freedreno_blitter.h"
 #include "freedreno_draw.h"
 #include "freedreno_context.h"
+#include "freedreno_fence.h"
 #include "freedreno_state.h"
 #include "freedreno_resource.h"
 #include "freedreno_query_acc.h"
@@ -61,22 +62,24 @@ static void
 fd_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
 {
        struct fd_context *ctx = fd_context(pctx);
-       struct fd_batch *batch = ctx->batch;
+       struct fd_batch *batch = fd_context_batch(ctx);
        struct pipe_framebuffer_state *pfb = &batch->framebuffer;
-       struct pipe_scissor_state *scissor = fd_context_get_scissor(ctx);
-       unsigned i, prims, buffers = 0;
+       unsigned i, prims, buffers = 0, restore_buffers = 0;
+
+       /* for debugging problems with indirect draw, it is convenient
+        * to be able to emulate it, to determine if game is feeding us
+        * bogus data:
+        */
+       if (info->indirect && (fd_mesa_debug & FD_DBG_NOINDR)) {
+               util_draw_indirect(pctx, info);
+               return;
+       }
 
        if (!info->count_from_stream_output && !info->indirect &&
            !info->primitive_restart &&
            !u_trim_pipe_prim(info->mode, (unsigned*)&info->count))
                return;
 
-       /* if we supported transform feedback, we'd have to disable this: */
-       if (((scissor->maxx - scissor->minx) *
-                       (scissor->maxy - scissor->miny)) == 0) {
-               return;
-       }
-
        /* TODO: push down the region versions into the tiles */
        if (!fd_render_condition_check(pctx))
                return;
@@ -90,6 +93,8 @@ fd_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
                return;
        }
 
+       fd_fence_ref(&ctx->last_fence, NULL);
+
        /* Upload a user index buffer. */
        struct pipe_resource *indexbuf = NULL;
        unsigned index_offset = 0;
@@ -126,16 +131,32 @@ fd_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
 
        mtx_lock(&ctx->screen->lock);
 
-       if (fd_depth_enabled(ctx)) {
-               buffers |= FD_BUFFER_DEPTH;
-               resource_written(batch, pfb->zsbuf->texture);
-               batch->gmem_reason |= FD_GMEM_DEPTH_ENABLED;
-       }
+       if (ctx->dirty & (FD_DIRTY_FRAMEBUFFER | FD_DIRTY_ZSA)) {
+               if (fd_depth_enabled(ctx)) {
+                       if (fd_resource(pfb->zsbuf->texture)->valid) {
+                               restore_buffers |= FD_BUFFER_DEPTH;
+                       } else {
+                               batch->invalidated |= FD_BUFFER_DEPTH;
+                       }
+                       batch->gmem_reason |= FD_GMEM_DEPTH_ENABLED;
+                       if (fd_depth_write_enabled(ctx)) {
+                               buffers |= FD_BUFFER_DEPTH;
+                               resource_written(batch, pfb->zsbuf->texture);
+                       } else {
+                               resource_read(batch, pfb->zsbuf->texture);
+                       }
+               }
 
-       if (fd_stencil_enabled(ctx)) {
-               buffers |= FD_BUFFER_STENCIL;
-               resource_written(batch, pfb->zsbuf->texture);
-               batch->gmem_reason |= FD_GMEM_STENCIL_ENABLED;
+               if (fd_stencil_enabled(ctx)) {
+                       if (fd_resource(pfb->zsbuf->texture)->valid) {
+                               restore_buffers |= FD_BUFFER_STENCIL;
+                       } else {
+                               batch->invalidated |= FD_BUFFER_STENCIL;
+                       }
+                       batch->gmem_reason |= FD_GMEM_STENCIL_ENABLED;
+                       buffers |= FD_BUFFER_STENCIL;
+                       resource_written(batch, pfb->zsbuf->texture);
+               }
        }
 
        if (fd_logicop_enabled(ctx))
@@ -149,46 +170,82 @@ fd_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
 
                surf = pfb->cbufs[i]->texture;
 
-               resource_written(batch, surf);
-               buffers |= PIPE_CLEAR_COLOR0 << i;
+               if (fd_resource(surf)->valid) {
+                       restore_buffers |= PIPE_CLEAR_COLOR0 << i;
+               } else {
+                       batch->invalidated |= PIPE_CLEAR_COLOR0 << i;
+               }
 
-               if (surf->nr_samples > 1)
-                       batch->gmem_reason |= FD_GMEM_MSAA_ENABLED;
+               buffers |= PIPE_CLEAR_COLOR0 << i;
 
                if (fd_blend_enabled(ctx, i))
                        batch->gmem_reason |= FD_GMEM_BLEND_ENABLED;
+
+               if (ctx->dirty & FD_DIRTY_FRAMEBUFFER)
+                       resource_written(batch, pfb->cbufs[i]->texture);
        }
 
        /* Mark SSBOs as being written.. we don't actually know which ones are
         * read vs written, so just assume the worst
         */
-       foreach_bit(i, ctx->shaderbuf[PIPE_SHADER_FRAGMENT].enabled_mask)
-               resource_read(batch, ctx->shaderbuf[PIPE_SHADER_FRAGMENT].sb[i].buffer);
+       if (ctx->dirty_shader[PIPE_SHADER_FRAGMENT] & FD_DIRTY_SHADER_SSBO) {
+               foreach_bit(i, ctx->shaderbuf[PIPE_SHADER_FRAGMENT].enabled_mask)
+                               resource_written(batch, ctx->shaderbuf[PIPE_SHADER_FRAGMENT].sb[i].buffer);
+       }
+
+       if (ctx->dirty_shader[PIPE_SHADER_FRAGMENT] & FD_DIRTY_SHADER_IMAGE) {
+               foreach_bit(i, ctx->shaderimg[PIPE_SHADER_FRAGMENT].enabled_mask) {
+                       struct pipe_image_view *img =
+                                       &ctx->shaderimg[PIPE_SHADER_FRAGMENT].si[i];
+                       if (img->access & PIPE_IMAGE_ACCESS_WRITE)
+                               resource_written(batch, img->resource);
+                       else
+                               resource_read(batch, img->resource);
+               }
+       }
 
-       foreach_bit(i, ctx->constbuf[PIPE_SHADER_VERTEX].enabled_mask)
-               resource_read(batch, ctx->constbuf[PIPE_SHADER_VERTEX].cb[i].buffer);
-       foreach_bit(i, ctx->constbuf[PIPE_SHADER_FRAGMENT].enabled_mask)
-               resource_read(batch, ctx->constbuf[PIPE_SHADER_FRAGMENT].cb[i].buffer);
+       if (ctx->dirty_shader[PIPE_SHADER_VERTEX] & FD_DIRTY_SHADER_CONST) {
+               foreach_bit(i, ctx->constbuf[PIPE_SHADER_VERTEX].enabled_mask)
+                       resource_read(batch, ctx->constbuf[PIPE_SHADER_VERTEX].cb[i].buffer);
+       }
+
+       if (ctx->dirty_shader[PIPE_SHADER_FRAGMENT] & FD_DIRTY_SHADER_CONST) {
+               foreach_bit(i, ctx->constbuf[PIPE_SHADER_FRAGMENT].enabled_mask)
+                       resource_read(batch, ctx->constbuf[PIPE_SHADER_FRAGMENT].cb[i].buffer);
+       }
 
        /* Mark VBOs as being read */
-       foreach_bit(i, ctx->vtx.vertexbuf.enabled_mask) {
-               assert(!ctx->vtx.vertexbuf.vb[i].is_user_buffer);
-               resource_read(batch, ctx->vtx.vertexbuf.vb[i].buffer.resource);
+       if (ctx->dirty & FD_DIRTY_VTXBUF) {
+               foreach_bit(i, ctx->vtx.vertexbuf.enabled_mask) {
+                       assert(!ctx->vtx.vertexbuf.vb[i].is_user_buffer);
+                       resource_read(batch, ctx->vtx.vertexbuf.vb[i].buffer.resource);
+               }
        }
 
        /* Mark index buffer as being read */
        resource_read(batch, indexbuf);
 
+       /* Mark indirect draw buffer as being read */
+       if (info->indirect)
+               resource_read(batch, info->indirect->buffer);
+
        /* Mark textures as being read */
-       foreach_bit(i, ctx->tex[PIPE_SHADER_VERTEX].valid_textures)
-               resource_read(batch, ctx->tex[PIPE_SHADER_VERTEX].textures[i]->texture);
-       foreach_bit(i, ctx->tex[PIPE_SHADER_FRAGMENT].valid_textures)
-               resource_read(batch, ctx->tex[PIPE_SHADER_FRAGMENT].textures[i]->texture);
+       if (ctx->dirty_shader[PIPE_SHADER_VERTEX] & FD_DIRTY_SHADER_TEX) {
+               foreach_bit(i, ctx->tex[PIPE_SHADER_VERTEX].valid_textures)
+                       resource_read(batch, ctx->tex[PIPE_SHADER_VERTEX].textures[i]->texture);
+       }
+
+       if (ctx->dirty_shader[PIPE_SHADER_FRAGMENT] & FD_DIRTY_SHADER_TEX) {
+               foreach_bit(i, ctx->tex[PIPE_SHADER_FRAGMENT].valid_textures)
+                       resource_read(batch, ctx->tex[PIPE_SHADER_FRAGMENT].textures[i]->texture);
+       }
 
        /* Mark streamout buffers as being written.. */
-       for (i = 0; i < ctx->streamout.num_targets; i++)
-               if (ctx->streamout.targets[i])
-                       resource_written(batch, ctx->streamout.targets[i]->buffer);
+       if (ctx->dirty & FD_DIRTY_STREAMOUT) {
+               for (i = 0; i < ctx->streamout.num_targets; i++)
+                       if (ctx->streamout.targets[i])
+                               resource_written(batch, ctx->streamout.targets[i]->buffer);
+       }
 
        resource_written(batch, batch->query_buf);
 
@@ -214,7 +271,7 @@ fd_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
        ctx->stats.prims_generated += prims;
 
        /* any buffers that haven't been cleared yet, we need to restore: */
-       batch->restore |= buffers & (FD_BUFFER_ALL & ~batch->cleared);
+       batch->restore |= restore_buffers & (FD_BUFFER_ALL & ~batch->invalidated);
        /* and any buffers used, need to be resolved: */
        batch->resolve |= buffers;
 
@@ -226,6 +283,8 @@ fd_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
        if (ctx->draw_vbo(ctx, info, index_offset))
                batch->needs_flush = true;
 
+       batch->num_vertices += info->count * info->instance_count;
+
        for (i = 0; i < ctx->streamout.num_targets; i++)
                ctx->streamout.offsets[i] += info->count;
 
@@ -238,89 +297,13 @@ fd_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
                pipe_resource_reference(&indexbuf, NULL);
 }
 
-/* Generic clear implementation (partially) using u_blitter: */
-static void
-fd_blitter_clear(struct pipe_context *pctx, unsigned buffers,
-               const union pipe_color_union *color, double depth, unsigned stencil)
-{
-       struct fd_context *ctx = fd_context(pctx);
-       struct pipe_framebuffer_state *pfb = &ctx->batch->framebuffer;
-       struct blitter_context *blitter = ctx->blitter;
-
-       fd_blitter_pipe_begin(ctx, false, true, FD_STAGE_CLEAR);
-
-       util_blitter_common_clear_setup(blitter, pfb->width, pfb->height,
-                       buffers, NULL, NULL);
-
-       struct pipe_stencil_ref sr = {
-               .ref_value = { stencil & 0xff }
-       };
-       pctx->set_stencil_ref(pctx, &sr);
-
-       struct pipe_constant_buffer cb = {
-               .buffer_size = 16,
-               .user_buffer = &color->ui,
-       };
-       pctx->set_constant_buffer(pctx, PIPE_SHADER_FRAGMENT, 0, &cb);
-
-       if (!ctx->clear_rs_state) {
-               const struct pipe_rasterizer_state tmpl = {
-                       .cull_face = PIPE_FACE_NONE,
-                       .half_pixel_center = 1,
-                       .bottom_edge_rule = 1,
-                       .flatshade = 1,
-                       .depth_clip = 1,
-               };
-               ctx->clear_rs_state = pctx->create_rasterizer_state(pctx, &tmpl);
-       }
-       pctx->bind_rasterizer_state(pctx, ctx->clear_rs_state);
-
-       struct pipe_viewport_state vp = {
-               .scale     = { 0.5f * pfb->width, -0.5f * pfb->height, depth },
-               .translate = { 0.5f * pfb->width,  0.5f * pfb->height, 0.0f },
-       };
-       pctx->set_viewport_states(pctx, 0, 1, &vp);
-
-       pctx->bind_vertex_elements_state(pctx, ctx->solid_vbuf_state.vtx);
-       pctx->set_vertex_buffers(pctx, blitter->vb_slot, 1,
-                       &ctx->solid_vbuf_state.vertexbuf.vb[0]);
-       pctx->set_stream_output_targets(pctx, 0, NULL, NULL);
-       pctx->bind_vs_state(pctx, ctx->solid_prog.vp);
-       pctx->bind_fs_state(pctx, ctx->solid_prog.fp);
-
-       struct pipe_draw_info info = {
-               .mode = PIPE_PRIM_MAX,    /* maps to DI_PT_RECTLIST */
-               .count = 2,
-               .max_index = 1,
-               .instance_count = 1,
-       };
-       ctx->draw_vbo(ctx, &info, 0);
-
-       util_blitter_restore_constant_buffer_state(blitter);
-       util_blitter_restore_vertex_states(blitter);
-       util_blitter_restore_fragment_states(blitter);
-       util_blitter_restore_textures(blitter);
-       util_blitter_restore_fb_state(blitter);
-       util_blitter_restore_render_cond(blitter);
-       util_blitter_unset_running_flag(blitter);
-
-       fd_blitter_pipe_end(ctx);
-}
-
-/* TODO figure out how to make better use of existing state mechanism
- * for clear (and possibly gmem->mem / mem->gmem) so we can (a) keep
- * track of what state really actually changes, and (b) reduce the code
- * in the a2xx/a3xx parts.
- */
-
 static void
 fd_clear(struct pipe_context *pctx, unsigned buffers,
                const union pipe_color_union *color, double depth, unsigned stencil)
 {
        struct fd_context *ctx = fd_context(pctx);
-       struct fd_batch *batch = ctx->batch;
+       struct fd_batch *batch = fd_context_batch(ctx);
        struct pipe_framebuffer_state *pfb = &batch->framebuffer;
-       struct pipe_scissor_state *scissor = fd_context_get_scissor(ctx);
        unsigned cleared_buffers;
        int i;
 
@@ -328,11 +311,21 @@ fd_clear(struct pipe_context *pctx, unsigned buffers,
        if (!fd_render_condition_check(pctx))
                return;
 
+       fd_fence_ref(&ctx->last_fence, NULL);
+
        if (ctx->in_blit) {
                fd_batch_reset(batch);
                fd_context_all_dirty(ctx);
        }
 
+       /* pctx->clear() is only for full-surface clears, so scissor is
+        * equivalent to having GL_SCISSOR_TEST disabled:
+        */
+       batch->max_scissor.minx = 0;
+       batch->max_scissor.miny = 0;
+       batch->max_scissor.maxx = pfb->width;
+       batch->max_scissor.maxy = pfb->height;
+
        /* 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
@@ -341,19 +334,9 @@ fd_clear(struct pipe_context *pctx, unsigned buffers,
         * the depth buffer, etc)
         */
        cleared_buffers = buffers & (FD_BUFFER_ALL & ~batch->restore);
+       batch->cleared |= buffers;
+       batch->invalidated |= cleared_buffers;
 
-       /* do we have full-screen scissor? */
-       if (!memcmp(scissor, &ctx->disabled_scissor, sizeof(*scissor))) {
-               batch->cleared |= cleared_buffers;
-       } else {
-               batch->partial_cleared |= cleared_buffers;
-               if (cleared_buffers & PIPE_CLEAR_COLOR)
-                       batch->cleared_scissor.color = *scissor;
-               if (cleared_buffers & PIPE_CLEAR_DEPTH)
-                       batch->cleared_scissor.depth = *scissor;
-               if (cleared_buffers & PIPE_CLEAR_STENCIL)
-                       batch->cleared_scissor.stencil = *scissor;
-       }
        batch->resolve |= buffers;
        batch->needs_flush = true;
 
@@ -400,6 +383,8 @@ fd_clear(struct pipe_context *pctx, unsigned buffers,
        if (fallback) {
                fd_blitter_clear(pctx, buffers, color, depth, stencil);
        }
+
+       fd_batch_check_size(batch);
 }
 
 static void
@@ -428,15 +413,10 @@ fd_launch_grid(struct pipe_context *pctx, const struct pipe_grid_info *info)
        struct fd_batch *batch, *save_batch = NULL;
        unsigned i;
 
-       /* TODO maybe we don't want to allocate and flush a batch each time?
-        * We could use a special bogus (ie. won't match any fb state) key
-        * in the batch-case for compute shaders, and rely on the rest of
-        * the dependency tracking mechanism to tell us when the compute
-        * batch needs to be flushed?
-        */
-       batch = fd_bc_alloc_batch(&ctx->screen->batch_cache, ctx);
+       batch = fd_bc_alloc_batch(&ctx->screen->batch_cache, ctx, true);
        fd_batch_reference(&save_batch, ctx->batch);
        fd_batch_reference(&ctx->batch, batch);
+       fd_context_all_dirty(ctx);
 
        mtx_lock(&ctx->screen->lock);
 
@@ -444,7 +424,16 @@ fd_launch_grid(struct pipe_context *pctx, const struct pipe_grid_info *info)
         * read vs written, so just assume the worst
         */
        foreach_bit(i, ctx->shaderbuf[PIPE_SHADER_COMPUTE].enabled_mask)
-               resource_read(batch, ctx->shaderbuf[PIPE_SHADER_COMPUTE].sb[i].buffer);
+               resource_written(batch, ctx->shaderbuf[PIPE_SHADER_COMPUTE].sb[i].buffer);
+
+       foreach_bit(i, ctx->shaderimg[PIPE_SHADER_COMPUTE].enabled_mask) {
+               struct pipe_image_view *img =
+                       &ctx->shaderimg[PIPE_SHADER_COMPUTE].si[i];
+               if (img->access & PIPE_IMAGE_ACCESS_WRITE)
+                       resource_written(batch, img->resource);
+               else
+                       resource_read(batch, img->resource);
+       }
 
        /* UBO's are read */
        foreach_bit(i, ctx->constbuf[PIPE_SHADER_COMPUTE].enabled_mask)
@@ -454,14 +443,26 @@ fd_launch_grid(struct pipe_context *pctx, const struct pipe_grid_info *info)
        foreach_bit(i, ctx->tex[PIPE_SHADER_COMPUTE].valid_textures)
                resource_read(batch, ctx->tex[PIPE_SHADER_COMPUTE].textures[i]->texture);
 
+       /* For global buffers, we don't really know if read or written, so assume
+        * the worst:
+        */
+       foreach_bit(i, ctx->global_bindings.enabled_mask)
+               resource_written(batch, ctx->global_bindings.buf[i]);
+
+       if (info->indirect)
+               resource_read(batch, info->indirect);
+
        mtx_unlock(&ctx->screen->lock);
 
+       batch->needs_flush = true;
        ctx->launch_grid(ctx, info);
 
-       fd_gmem_flush_compute(batch);
+       fd_batch_flush(batch, false);
 
        fd_batch_reference(&ctx->batch, save_batch);
+       fd_context_all_dirty(ctx);
        fd_batch_reference(&save_batch, NULL);
+       fd_batch_reference(&batch, NULL);
 }
 
 void