radeonsi: Set PIPE_SHADER_CAP_MAX_SHADER_IMAGES
[mesa.git] / src / gallium / drivers / radeonsi / si_state.c
index 8190913df8a081bc032b52753f4fdf2fbbee4351..b9bdd47c4967301587f6275fa9e802c6bb9f0914 100644 (file)
@@ -34,6 +34,7 @@
 #include "util/u_format_s3tc.h"
 #include "util/u_memory.h"
 #include "util/u_pstipple.h"
+#include "util/u_resource.h"
 
 /* Initialize an external atom (owned by ../radeon). */
 static void
@@ -2796,7 +2797,7 @@ static void si_set_min_samples(struct pipe_context *ctx, unsigned min_samples)
  * Build the sampler view descriptor for a buffer texture.
  * @param state 256-bit descriptor; only the high 128 bits are filled in
  */
-static void
+void
 si_make_buffer_descriptor(struct si_screen *screen, struct r600_resource *buf,
                          enum pipe_format format,
                          unsigned first_element, unsigned last_element,
@@ -2837,9 +2838,10 @@ si_make_buffer_descriptor(struct si_screen *screen, struct r600_resource *buf,
 /**
  * Build the sampler view descriptor for a texture.
  */
-static void
+void
 si_make_texture_descriptor(struct si_screen *screen,
                           struct r600_texture *tex,
+                          bool sampler,
                           enum pipe_texture_target target,
                           enum pipe_format pipe_format,
                           const unsigned char state_swizzle[4],
@@ -2854,7 +2856,7 @@ si_make_texture_descriptor(struct si_screen *screen,
        const struct util_format_description *desc;
        unsigned char swizzle[4];
        int first_non_void;
-       unsigned num_format, data_format;
+       unsigned num_format, data_format, type;
        uint32_t pitch;
        uint64_t va;
 
@@ -2972,12 +2974,29 @@ si_make_texture_descriptor(struct si_screen *screen,
                data_format = 0;
        }
 
-       if (res->target == PIPE_TEXTURE_1D_ARRAY) {
+       if (!sampler &&
+           (res->target == PIPE_TEXTURE_CUBE ||
+            res->target == PIPE_TEXTURE_CUBE_ARRAY ||
+            res->target == PIPE_TEXTURE_3D)) {
+               /* For the purpose of shader images, treat cube maps and 3D
+                * textures as 2D arrays. For 3D textures, the address
+                * calculations for mipmaps are different, so we rely on the
+                * caller to effectively disable mipmaps.
+                */
+               type = V_008F1C_SQ_RSRC_IMG_2D_ARRAY;
+
+               assert(res->target != PIPE_TEXTURE_3D || (first_level == 0 && last_level == 0));
+       } else {
+               type = si_tex_dim(res->target, target, res->nr_samples);
+       }
+
+       if (type == V_008F1C_SQ_RSRC_IMG_1D_ARRAY) {
                height = 1;
                depth = res->array_size;
-       } else if (res->target == PIPE_TEXTURE_2D_ARRAY) {
-               depth = res->array_size;
-       } else if (res->target == PIPE_TEXTURE_CUBE_ARRAY)
+       } else if (type == V_008F1C_SQ_RSRC_IMG_2D_ARRAY) {
+               if (sampler || res->target != PIPE_TEXTURE_3D)
+                       depth = res->array_size;
+       } else if (type == V_008F1C_SQ_RSRC_IMG_CUBE)
                depth = res->array_size / 6;
 
        pitch = surflevel[base_level].nblk_x * util_format_get_blockwidth(pipe_format);
@@ -3000,7 +3019,7 @@ si_make_texture_descriptor(struct si_screen *screen,
                                        last_level) |
                    S_008F1C_TILING_INDEX(si_tile_mode_index(tex, base_level, false)) |
                    S_008F1C_POW2_PAD(res->last_level > 0) |
-                   S_008F1C_TYPE(si_tex_dim(res->target, target, res->nr_samples)));
+                   S_008F1C_TYPE(type));
        state[4] = (S_008F20_DEPTH(depth - 1) | S_008F20_PITCH(pitch - 1));
        state[5] = (S_008F24_BASE_ARRAY(first_layer) |
                    S_008F24_LAST_ARRAY(last_layer));
@@ -3154,7 +3173,7 @@ si_create_sampler_view_custom(struct pipe_context *ctx,
            state->target == PIPE_TEXTURE_CUBE)
                last_layer = state->u.tex.first_layer;
 
-       si_make_texture_descriptor(sctx->screen, tmp, state->target,
+       si_make_texture_descriptor(sctx->screen, tmp, true, state->target,
                                   state->format, state_swizzle,
                                   base_level, first_level, last_level,
                                   state->u.tex.first_layer, last_layer,
@@ -3502,6 +3521,42 @@ static void si_texture_barrier(struct pipe_context *ctx)
                         SI_CONTEXT_FLUSH_AND_INV_CB;
 }
 
+static void si_memory_barrier(struct pipe_context *ctx, unsigned flags)
+{
+       struct si_context *sctx = (struct si_context *)ctx;
+
+       /* Subsequent commands must wait for all shader invocations to
+        * complete. */
+       sctx->b.flags |= SI_CONTEXT_PS_PARTIAL_FLUSH;
+
+       if (flags & PIPE_BARRIER_CONSTANT_BUFFER)
+               sctx->b.flags |= SI_CONTEXT_INV_SMEM_L1 |
+                                SI_CONTEXT_INV_VMEM_L1;
+
+       if (flags & (PIPE_BARRIER_VERTEX_BUFFER |
+                    PIPE_BARRIER_SHADER_BUFFER |
+                    PIPE_BARRIER_TEXTURE |
+                    PIPE_BARRIER_IMAGE)) {
+               /* As far as I can tell, L1 contents are written back to L2
+                * automatically at end of shader, but the contents of other
+                * L1 caches might still be stale. */
+               sctx->b.flags |= SI_CONTEXT_INV_VMEM_L1;
+       }
+
+       if (flags & PIPE_BARRIER_FRAMEBUFFER)
+               sctx->b.flags |= SI_CONTEXT_FLUSH_AND_INV_FRAMEBUFFER;
+
+       if (flags & (PIPE_BARRIER_MAPPED_BUFFER |
+                    PIPE_BARRIER_FRAMEBUFFER)) {
+               /* Not sure if INV_GLOBAL_L2 is the best thing here.
+                *
+                * We need to make sure that TC L1 & L2 are written back to
+                * memory, because neither CPU accesses nor CB fetches consider
+                * TC, but there's no need to invalidate any TC cache lines. */
+               sctx->b.flags |= SI_CONTEXT_INV_GLOBAL_L2;
+       }
+}
+
 static void *si_create_blend_custom(struct si_context *sctx, unsigned mode)
 {
        struct pipe_blend_state blend;
@@ -3582,6 +3637,7 @@ void si_init_state_functions(struct si_context *sctx)
        sctx->b.b.set_index_buffer = si_set_index_buffer;
 
        sctx->b.b.texture_barrier = si_texture_barrier;
+       sctx->b.b.memory_barrier = si_memory_barrier;
        sctx->b.b.set_polygon_stipple = si_set_polygon_stipple;
        sctx->b.b.set_min_samples = si_set_min_samples;
        sctx->b.b.set_tess_state = si_set_tess_state;
@@ -3600,6 +3656,69 @@ void si_init_state_functions(struct si_context *sctx)
        si_init_config(sctx);
 }
 
+static void si_query_opaque_metadata(struct r600_common_screen *rscreen,
+                                    struct r600_texture *rtex,
+                                    struct radeon_bo_metadata *md)
+{
+       struct si_screen *sscreen = (struct si_screen*)rscreen;
+       struct pipe_resource *res = &rtex->resource.b.b;
+       static const unsigned char swizzle[] = {
+               PIPE_SWIZZLE_RED,
+               PIPE_SWIZZLE_GREEN,
+               PIPE_SWIZZLE_BLUE,
+               PIPE_SWIZZLE_ALPHA
+       };
+       uint32_t desc[8], i;
+       bool is_array = util_resource_is_array_texture(res);
+
+       /* DRM 2.x.x doesn't support this. */
+       if (rscreen->info.drm_major != 3)
+               return;
+
+       assert(rtex->fmask.size == 0);
+
+       /* Metadata image format format version 1:
+        * [0] = 1 (metadata format identifier)
+        * [1] = (VENDOR_ID << 16) | PCI_ID
+        * [2:9] = image descriptor for the whole resource
+        *         [2] is always 0, because the base address is cleared
+        *         [9] is the DCC offset bits [39:8] from the beginning of
+        *             the buffer
+        * [10:10+LAST_LEVEL] = mipmap level offset bits [39:8] for each level
+        */
+
+       md->metadata[0] = 1; /* metadata image format version 1 */
+
+       /* TILE_MODE_INDEX is ambiguous without a PCI ID. */
+       md->metadata[1] = (ATI_VENDOR_ID << 16) | rscreen->info.pci_id;
+
+       si_make_texture_descriptor(sscreen, rtex, true,
+                                  res->target, res->format,
+                                  swizzle, 0, 0, res->last_level, 0,
+                                  is_array ? res->array_size - 1 : 0,
+                                  res->width0, res->height0, res->depth0,
+                                  desc, NULL);
+
+       /* Clear the base address and set the relative DCC offset. */
+       desc[0] = 0;
+       desc[1] &= C_008F14_BASE_ADDRESS_HI;
+       desc[7] = rtex->dcc_offset >> 8;
+
+       /* Dwords [2:9] contain the image descriptor. */
+       memcpy(&md->metadata[2], desc, sizeof(desc));
+
+       /* Dwords [10:..] contain the mipmap level offsets. */
+       for (i = 0; i <= res->last_level; i++)
+               md->metadata[10+i] = rtex->surface.level[i].offset >> 8;
+
+       md->size_metadata = (11 + res->last_level) * 4;
+}
+
+void si_init_screen_state_functions(struct si_screen *sscreen)
+{
+       sscreen->b.query_opaque_metadata = si_query_opaque_metadata;
+}
+
 static void
 si_write_harvested_raster_configs(struct si_context *sctx,
                                  struct si_pm4_state *pm4,