vc4: Cache texture p0/p1 setup for the sampler view.

In exchange for a bit of space and computation in CSO setup, we cut
vc4_uniform.c (draw time) code size by 4.8%.

3 files changed, 43 insertions, 28 deletions

diff --git a/src/gallium/drivers/vc4/vc4_context.h b/src/gallium/drivers/vc4/vc4_context.h
index d5d6be16f6e..be7df1bd347 100644
--- a/src/gallium/drivers/vc4/vc4_context.h
+++ b/src/gallium/drivers/vc4/vc4_context.h
@@ -69,6 +69,12 @@
 #define VC4_DIRTY_UNCOMPILED_FS (1 << 22)
 #define VC4_DIRTY_COMPILED_FS   (1 << 24)
 
+struct vc4_sampler_view {
+        struct pipe_sampler_view base;
+        uint32_t texture_p0;
+        uint32_t texture_p1;
+};
+
 struct vc4_texture_stateobj {
         struct pipe_sampler_view *textures[PIPE_MAX_SAMPLERS];
         unsigned num_textures;
@@ -326,6 +332,12 @@ vc4_context(struct pipe_context *pcontext)
         return (struct vc4_context *)pcontext;
 }
 
+static inline struct vc4_sampler_view *
+vc4_sampler_view(struct pipe_sampler_view *psview)
+{
+        return (struct vc4_sampler_view *)psview;
+}
+
 struct pipe_context *vc4_context_create(struct pipe_screen *pscreen,
                                         void *priv);
 void vc4_draw_init(struct pipe_context *pctx);
diff --git a/src/gallium/drivers/vc4/vc4_state.c b/src/gallium/drivers/vc4/vc4_state.c
index e0ce4aee779..46852b6628c 100644
--- a/src/gallium/drivers/vc4/vc4_state.c
+++ b/src/gallium/drivers/vc4/vc4_state.c
@@ -499,13 +499,13 @@ static struct pipe_sampler_view *
 vc4_create_sampler_view(struct pipe_context *pctx, struct pipe_resource *prsc,
                         const struct pipe_sampler_view *cso)
 {
-        struct pipe_sampler_view *so = malloc(sizeof(*so));
+        struct vc4_sampler_view *so = malloc(sizeof(*so));
         struct vc4_resource *rsc = vc4_resource(prsc);
 
         if (!so)
                 return NULL;
 
-        *so = *cso;
+        so->base = *cso;
 
         pipe_reference(NULL, &prsc->reference);
 
@@ -516,18 +516,19 @@ vc4_create_sampler_view(struct pipe_context *pctx, struct pipe_resource *prsc,
          * Also, Raspberry Pi doesn't support sampling from raster textures,
          * so we also have to copy to a temporary then.
          */
-        if (so->u.tex.first_level ||
+        if (cso->u.tex.first_level ||
             rsc->vc4_format == VC4_TEXTURE_TYPE_RGBA32R) {
                 struct vc4_resource *shadow_parent = vc4_resource(prsc);
                 struct pipe_resource tmpl = shadow_parent->base.b;
                 struct vc4_resource *clone;
 
                 tmpl.bind = PIPE_BIND_SAMPLER_VIEW | PIPE_BIND_RENDER_TARGET;
-                tmpl.width0 = u_minify(tmpl.width0, so->u.tex.first_level);
-                tmpl.height0 = u_minify(tmpl.height0, so->u.tex.first_level);
-                tmpl.last_level = so->u.tex.last_level - so->u.tex.first_level;
+                tmpl.width0 = u_minify(tmpl.width0, cso->u.tex.first_level);
+                tmpl.height0 = u_minify(tmpl.height0, cso->u.tex.first_level);
+                tmpl.last_level = cso->u.tex.last_level - cso->u.tex.first_level;
 
                 prsc = vc4_resource_create(pctx->screen, &tmpl);
+                rsc = vc4_resource(prsc);
                 clone = vc4_resource(prsc);
                 clone->shadow_parent = &shadow_parent->base.b;
                 /* Flag it as needing update of the contents from the parent. */
@@ -535,11 +536,23 @@ vc4_create_sampler_view(struct pipe_context *pctx, struct pipe_resource *prsc,
 
                 assert(clone->vc4_format != VC4_TEXTURE_TYPE_RGBA32R);
         }
-        so->texture = prsc;
-        so->reference.count = 1;
-        so->context = pctx;
-
-        return so;
+        so->base.texture = prsc;
+        so->base.reference.count = 1;
+        so->base.context = pctx;
+
+        so->texture_p0 =
+                (VC4_SET_FIELD(rsc->slices[0].offset >> 12, VC4_TEX_P0_OFFSET) |
+                 VC4_SET_FIELD(rsc->vc4_format & 15, VC4_TEX_P0_TYPE) |
+                 VC4_SET_FIELD(cso->u.tex.last_level -
+                               cso->u.tex.first_level, VC4_TEX_P0_MIPLVLS) |
+                 VC4_SET_FIELD(cso->target == PIPE_TEXTURE_CUBE,
+                               VC4_TEX_P0_CMMODE));
+        so->texture_p1 =
+                (VC4_SET_FIELD(rsc->vc4_format >> 4, VC4_TEX_P1_TYPE4) |
+                 VC4_SET_FIELD(prsc->height0 & 2047, VC4_TEX_P1_HEIGHT) |
+                 VC4_SET_FIELD(prsc->width0 & 2047, VC4_TEX_P1_WIDTH));
+
+        return &so->base;
 }
 
 static void
diff --git a/src/gallium/drivers/vc4/vc4_uniforms.c b/src/gallium/drivers/vc4/vc4_uniforms.c
index e8fca3d0cbd..e8e9d6b0119 100644
--- a/src/gallium/drivers/vc4/vc4_uniforms.c
+++ b/src/gallium/drivers/vc4/vc4_uniforms.c
@@ -53,16 +53,11 @@ write_texture_p0(struct vc4_context *vc4,
                  struct vc4_texture_stateobj *texstate,
                  uint32_t unit)
 {
-        struct pipe_sampler_view *texture = texstate->textures[unit];
-        struct vc4_resource *rsc = vc4_resource(texture->texture);
+        struct vc4_sampler_view *sview =
+                vc4_sampler_view(texstate->textures[unit]);
+        struct vc4_resource *rsc = vc4_resource(sview->base.texture);
 
-        cl_reloc(vc4, &vc4->uniforms, uniforms, rsc->bo,
-                 VC4_SET_FIELD(rsc->slices[0].offset >> 12, VC4_TEX_P0_OFFSET) |
-                 VC4_SET_FIELD(texture->u.tex.last_level -
-                               texture->u.tex.first_level, VC4_TEX_P0_MIPLVLS) |
-                 VC4_SET_FIELD(texture->target == PIPE_TEXTURE_CUBE,
-                               VC4_TEX_P0_CMMODE) |
-                 VC4_SET_FIELD(rsc->vc4_format & 15, VC4_TEX_P0_TYPE));
+        cl_reloc(vc4, &vc4->uniforms, uniforms, rsc->bo, sview->texture_p0);
 }
 
 static void
@@ -71,8 +66,8 @@ write_texture_p1(struct vc4_context *vc4,
                  struct vc4_texture_stateobj *texstate,
                  uint32_t unit)
 {
-        struct pipe_sampler_view *texture = texstate->textures[unit];
-        struct vc4_resource *rsc = vc4_resource(texture->texture);
+        struct vc4_sampler_view *sview =
+                vc4_sampler_view(texstate->textures[unit]);
         struct pipe_sampler_state *sampler = texstate->samplers[unit];
         static const uint8_t minfilter_map[6] = {
                 VC4_TEX_P1_MINFILT_NEAR_MIP_NEAR,
@@ -91,12 +86,7 @@ write_texture_p1(struct vc4_context *vc4,
                 (sampler->mag_img_filter == PIPE_TEX_MIPFILTER_NEAREST ||
                  sampler->min_img_filter == PIPE_TEX_MIPFILTER_NEAREST);
 
-        cl_aligned_u32(uniforms,
-               VC4_SET_FIELD(rsc->vc4_format >> 4, VC4_TEX_P1_TYPE4) |
-               VC4_SET_FIELD(texture->texture->height0 & 2047,
-                             VC4_TEX_P1_HEIGHT) |
-               VC4_SET_FIELD(texture->texture->width0 & 2047,
-                             VC4_TEX_P1_WIDTH) |
+        cl_aligned_u32(uniforms, sview->texture_p1 |
                VC4_SET_FIELD(magfilter_map[sampler->mag_img_filter],
                              VC4_TEX_P1_MAGFILT) |
                VC4_SET_FIELD(minfilter_map[sampler->min_mip_filter * 2 +