v3d: Add support for shader_image_load_store.

This is only exposed on V3D 4.1+, because we didn't have the TMU write
operations for images on 3.3 (To do GLES 3.1 there, you have to lower it
to SSBO load/stores, which is a problem to solve later).

5 files changed, 196 insertions, 2 deletions

diff --git a/src/gallium/drivers/v3d/v3d_context.h b/src/gallium/drivers/v3d/v3d_context.h
index 686ecfa8024..136d8506597 100644
--- a/src/gallium/drivers/v3d/v3d_context.h
+++ b/src/gallium/drivers/v3d/v3d_context.h
@@ -58,6 +58,7 @@ void v3d_job_add_bo(struct v3d_job *job, struct v3d_bo *bo);
 #define VC5_DIRTY_ZSA           (1 <<  2)
 #define VC5_DIRTY_FRAGTEX       (1 <<  3)
 #define VC5_DIRTY_VERTTEX       (1 <<  4)
+#define VC5_DIRTY_SHADER_IMAGE  (1 <<  5)
 
 #define VC5_DIRTY_BLEND_COLOR   (1 <<  7)
 #define VC5_DIRTY_STENCIL_REF   (1 <<  8)
@@ -222,6 +223,18 @@ enum v3d_ez_state {
         VC5_EZ_DISABLED,
 };
 
+struct v3d_image_view {
+        struct pipe_image_view base;
+        /* V3D 4.x texture shader state struct */
+        struct pipe_resource *tex_state;
+        uint32_t tex_state_offset;
+};
+
+struct v3d_shaderimg_stateobj {
+        struct v3d_image_view si[PIPE_MAX_SHADER_IMAGES];
+        uint32_t enabled_mask;
+};
+
 /**
  * A complete bin/render job.
  *
@@ -440,6 +453,7 @@ struct v3d_context {
         struct pipe_clip_state clip;
         struct pipe_viewport_state viewport;
         struct v3d_ssbo_stateobj ssbo[PIPE_SHADER_TYPES];
+        struct v3d_shaderimg_stateobj shaderimg[PIPE_SHADER_TYPES];
         struct v3d_constbuf_stateobj constbuf[PIPE_SHADER_TYPES];
         struct v3d_texture_stateobj tex[PIPE_SHADER_TYPES];
         struct v3d_vertexbuf_stateobj vertexbuf;
diff --git a/src/gallium/drivers/v3d/v3d_screen.c b/src/gallium/drivers/v3d/v3d_screen.c
index dd5eda87e8f..bee036a96d3 100644
--- a/src/gallium/drivers/v3d/v3d_screen.c
+++ b/src/gallium/drivers/v3d/v3d_screen.c
@@ -240,6 +240,8 @@ static int
 v3d_screen_get_shader_param(struct pipe_screen *pscreen, unsigned shader,
                            enum pipe_shader_cap param)
 {
+        struct v3d_screen *screen = v3d_screen(pscreen);
+
         if (shader != PIPE_SHADER_VERTEX &&
             shader != PIPE_SHADER_FRAGMENT) {
                 return 0;
@@ -298,12 +300,17 @@ v3d_screen_get_shader_param(struct pipe_screen *pscreen, unsigned shader,
                 return 1;
         case PIPE_SHADER_CAP_MAX_TEXTURE_SAMPLERS:
         case PIPE_SHADER_CAP_MAX_SAMPLER_VIEWS:
-        case PIPE_SHADER_CAP_MAX_SHADER_IMAGES:
                 return VC5_MAX_TEXTURE_SAMPLERS;
 
         case PIPE_SHADER_CAP_MAX_SHADER_BUFFERS:
                 return PIPE_MAX_SHADER_BUFFERS;
 
+        case PIPE_SHADER_CAP_MAX_SHADER_IMAGES:
+                if (screen->devinfo.ver < 41)
+                        return 0;
+                else
+                        return PIPE_MAX_SHADER_IMAGES;
+
         case PIPE_SHADER_CAP_PREFERRED_IR:
                 return PIPE_SHADER_IR_NIR;
         case PIPE_SHADER_CAP_SUPPORTED_IRS:
diff --git a/src/gallium/drivers/v3d/v3d_uniforms.c b/src/gallium/drivers/v3d/v3d_uniforms.c
index 5d907726349..5a3286013cd 100644
--- a/src/gallium/drivers/v3d/v3d_uniforms.c
+++ b/src/gallium/drivers/v3d/v3d_uniforms.c
@@ -71,6 +71,30 @@ get_texture_size(struct v3d_texture_stateobj *texstate,
         }
 }
 
+static uint32_t
+get_image_size(struct v3d_shaderimg_stateobj *shaderimg,
+               enum quniform_contents contents,
+               uint32_t data)
+{
+        struct v3d_image_view *image = &shaderimg->si[data];
+
+        switch (contents) {
+        case QUNIFORM_IMAGE_WIDTH:
+                return u_minify(image->base.resource->width0,
+                                image->base.u.tex.level);
+        case QUNIFORM_IMAGE_HEIGHT:
+                return u_minify(image->base.resource->height0,
+                                image->base.u.tex.level);
+        case QUNIFORM_IMAGE_DEPTH:
+                return u_minify(image->base.resource->depth0,
+                                image->base.u.tex.level);
+        case QUNIFORM_IMAGE_ARRAY_SIZE:
+                return image->base.resource->array_size;
+        default:
+                unreachable("Bad texture size field");
+        }
+}
+
 static struct v3d_bo *
 v3d_upload_ubo(struct v3d_context *v3d,
                struct v3d_compiled_shader *shader,
@@ -158,6 +182,27 @@ write_tmu_p0(struct v3d_job *job,
         v3d_job_add_bo(job, rsc->bo);
 }
 
+static void
+write_image_tmu_p0(struct v3d_job *job,
+                   struct v3d_cl_out **uniforms,
+                   struct v3d_shaderimg_stateobj *img,
+                   uint32_t data)
+{
+        /* Extract the image unit from the top bits, and the compiler's
+         * packed p0 from the bottom.
+         */
+        uint32_t unit = data >> 24;
+        uint32_t p0 = data & 0x00ffffff;
+
+        struct v3d_image_view *iview = &img->si[unit];
+        struct v3d_resource *rsc = v3d_resource(iview->base.resource);
+
+        cl_aligned_reloc(&job->indirect, uniforms,
+                         v3d_resource(iview->tex_state)->bo,
+                         iview->tex_state_offset | p0);
+        v3d_job_add_bo(job, rsc->bo);
+}
+
 /** Writes the V3D 4.x TMU configuration parameter 1. */
 static void
 write_tmu_p1(struct v3d_job *job,
@@ -232,6 +277,11 @@ v3d_write_uniforms(struct v3d_context *v3d, struct v3d_compiled_shader *shader,
                         write_tmu_p1(job, &uniforms, texstate, data);
                         break;
 
+                case QUNIFORM_IMAGE_TMU_CONFIG_P0:
+                        write_image_tmu_p0(job, &uniforms,
+                                           &v3d->shaderimg[stage], data);
+                        break;
+
                 case QUNIFORM_TEXTURE_CONFIG_P1:
                         write_texture_p1(job, &uniforms, texstate,
                                          data);
@@ -256,6 +306,16 @@ v3d_write_uniforms(struct v3d_context *v3d, struct v3d_compiled_shader *shader,
                                                         data));
                         break;
 
+                case QUNIFORM_IMAGE_WIDTH:
+                case QUNIFORM_IMAGE_HEIGHT:
+                case QUNIFORM_IMAGE_DEPTH:
+                case QUNIFORM_IMAGE_ARRAY_SIZE:
+                        cl_aligned_u32(&uniforms,
+                                       get_image_size(&v3d->shaderimg[stage],
+                                                      uinfo->contents[i],
+                                                      data));
+                        break;
+
                 case QUNIFORM_ALPHA_REF:
                         cl_aligned_f(&uniforms,
                                      v3d->zsa->base.alpha.ref_value);
@@ -382,6 +442,14 @@ v3d_set_shader_uniform_dirty_flags(struct v3d_compiled_shader *shader)
                         dirty |= VC5_DIRTY_SSBO;
                         break;
 
+                case QUNIFORM_IMAGE_TMU_CONFIG_P0:
+                case QUNIFORM_IMAGE_WIDTH:
+                case QUNIFORM_IMAGE_HEIGHT:
+                case QUNIFORM_IMAGE_DEPTH:
+                case QUNIFORM_IMAGE_ARRAY_SIZE:
+                        dirty |= VC5_DIRTY_SHADER_IMAGE;
+                        break;
+
                 case QUNIFORM_ALPHA_REF:
                         dirty |= VC5_DIRTY_ZSA;
                         break;
diff --git a/src/gallium/drivers/v3d/v3dx_draw.c b/src/gallium/drivers/v3d/v3dx_draw.c
index 7f111bbe75f..7ec687031d6 100644
--- a/src/gallium/drivers/v3d/v3dx_draw.c
+++ b/src/gallium/drivers/v3d/v3dx_draw.c
@@ -148,6 +148,13 @@ v3d_predraw_check_stage_inputs(struct pipe_context *pctx,
                 if (cb->buffer)
                         v3d_flush_jobs_writing_resource(v3d, cb->buffer);
         }
+
+        /* Flush writes to our image views */
+        foreach_bit(i, v3d->shaderimg[s].enabled_mask) {
+                struct v3d_image_view *view = &v3d->shaderimg[s].si[i];
+
+                v3d_flush_jobs_writing_resource(v3d, view->base.resource);
+        }
 }
 
 static void
@@ -487,6 +494,12 @@ v3d_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
                                                    v3d->ssbo[s].sb[i].buffer);
                         job->tmu_dirty_rcl = true;
                 }
+
+                foreach_bit(i, v3d->shaderimg[s].enabled_mask) {
+                        v3d_job_add_write_resource(job,
+                                                   v3d->shaderimg[s].si[i].base.resource);
+                        job->tmu_dirty_rcl = true;
+                }
         }
 
         /* Get space to emit our draw call into the BCL, using a branch to
diff --git a/src/gallium/drivers/v3d/v3dx_state.c b/src/gallium/drivers/v3d/v3dx_state.c
index 7ea5dabbe1f..5377f6173c3 100644
--- a/src/gallium/drivers/v3d/v3dx_state.c
+++ b/src/gallium/drivers/v3d/v3dx_state.c
@@ -884,7 +884,7 @@ v3d_create_sampler_view(struct pipe_context *pctx, struct pipe_resource *prsc,
                         tex.srgb = false;
                 } else {
                         tex.texture_type = v3d_get_tex_format(&screen->devinfo,
-                                                               cso->format);
+                                                              cso->format);
                 }
         };
 
@@ -1033,6 +1033,97 @@ v3d_set_shader_buffers(struct pipe_context *pctx,
         v3d->dirty |= VC5_DIRTY_SSBO;
 }
 
+static void
+v3d_create_image_view_texture_shader_state(struct v3d_context *v3d,
+                                           struct v3d_shaderimg_stateobj *so,
+                                           int img)
+{
+#if V3D_VERSION >= 40
+        struct v3d_image_view *iview = &so->si[img];
+
+        void *map;
+        u_upload_alloc(v3d->uploader, 0, cl_packet_length(TEXTURE_SHADER_STATE),
+                       32,
+                       &iview->tex_state_offset,
+                       &iview->tex_state,
+                       &map);
+
+        struct pipe_resource *prsc = iview->base.resource;
+
+        v3dx_pack(map, TEXTURE_SHADER_STATE, tex) {
+                v3d_setup_texture_shader_state(&tex, prsc,
+                                               iview->base.u.tex.level,
+                                               iview->base.u.tex.level,
+                                               iview->base.u.tex.first_layer,
+                                               iview->base.u.tex.last_layer);
+
+                tex.swizzle_r = translate_swizzle(PIPE_SWIZZLE_X);
+                tex.swizzle_g = translate_swizzle(PIPE_SWIZZLE_Y);
+                tex.swizzle_b = translate_swizzle(PIPE_SWIZZLE_Z);
+                tex.swizzle_a = translate_swizzle(PIPE_SWIZZLE_W);
+
+                tex.texture_type = v3d_get_tex_format(&v3d->screen->devinfo,
+                                                      iview->base.format);
+        };
+#else /* V3D_VERSION < 40 */
+        /* V3D 3.x doesn't use support shader image load/store operations on
+         * textures, so it would get lowered in the shader to general memory
+         * acceses.
+         */
+#endif
+}
+
+static void
+v3d_set_shader_images(struct pipe_context *pctx,
+                      enum pipe_shader_type shader,
+                      unsigned start, unsigned count,
+                      const struct pipe_image_view *images)
+{
+        struct v3d_context *v3d = v3d_context(pctx);
+        struct v3d_shaderimg_stateobj *so = &v3d->shaderimg[shader];
+
+        if (images) {
+                for (unsigned i = 0; i < count; i++) {
+                        unsigned n = i + start;
+                        struct v3d_image_view *iview = &so->si[n];
+
+                        if ((iview->base.resource == images[i].resource) &&
+                            (iview->base.format == images[i].format) &&
+                            (iview->base.access == images[i].access) &&
+                            !memcmp(&iview->base.u, &images[i].u,
+                                    sizeof(iview->base.u)))
+                                continue;
+
+                        util_copy_image_view(&iview->base, &images[i]);
+
+                        if (iview->base.resource) {
+                                so->enabled_mask |= 1 << n;
+                                v3d_create_image_view_texture_shader_state(v3d,
+                                                                           so,
+                                                                           n);
+                        } else {
+                                so->enabled_mask &= ~(1 << n);
+                                pipe_resource_reference(&iview->tex_state, NULL);
+                        }
+                }
+        } else {
+                for (unsigned i = 0; i < count; i++) {
+                        unsigned n = i + start;
+                        struct v3d_image_view *iview = &so->si[n];
+
+                        pipe_resource_reference(&iview->base.resource, NULL);
+                        pipe_resource_reference(&iview->tex_state, NULL);
+                }
+
+                if (count == 32)
+                        so->enabled_mask = 0;
+                else
+                        so->enabled_mask &= ~(((1 << count) - 1) << start);
+        }
+
+        v3d->dirty |= VC5_DIRTY_SHADER_IMAGE;
+}
+
 void
 v3dX(state_init)(struct pipe_context *pctx)
 {
@@ -1073,6 +1164,7 @@ v3dX(state_init)(struct pipe_context *pctx)
         pctx->set_sampler_views = v3d_set_sampler_views;
 
         pctx->set_shader_buffers = v3d_set_shader_buffers;
+        pctx->set_shader_images = v3d_set_shader_images;
 
         pctx->create_stream_output_target = v3d_create_stream_output_target;
         pctx->stream_output_target_destroy = v3d_stream_output_target_destroy;