st/nine: Cache constant buffer size

The shader constant buffer size with the
constant compaction code can vary depending
on the shader variant compiled (for example if
fog constants are required, etc).
Thus instead of using fixed size for the shader,
add in the variant cache the size required, pass it
to the context, and use this value.

Signed-off-by: Axel Davy <davyaxel0@gmail.com>

7 files changed, 52 insertions, 23 deletions

diff --git a/src/gallium/state_trackers/nine/nine_shader.h b/src/gallium/state_trackers/nine/nine_shader.h
index aa0c0ab1f7a..65dc2efd62a 100644
--- a/src/gallium/state_trackers/nine/nine_shader.h
+++ b/src/gallium/state_trackers/nine/nine_shader.h
@@ -119,16 +119,21 @@ struct nine_shader_variant
     struct nine_shader_variant *next;
     void *cso;
     unsigned *const_ranges;
+    unsigned const_used_size;
     uint64_t key;
 };
 
 static inline void *
-nine_shader_variant_get(struct nine_shader_variant *list, unsigned **const_ranges, uint64_t key)
+nine_shader_variant_get(struct nine_shader_variant *list,
+                        unsigned **const_ranges,
+                        unsigned *const_used_size,
+                        uint64_t key)
 {
     while (list->key != key && list->next)
         list = list->next;
     if (list->key == key) {
         *const_ranges = list->const_ranges;
+        *const_used_size = list->const_used_size;
         return list->cso;
     }
     return NULL;
@@ -136,7 +141,9 @@ nine_shader_variant_get(struct nine_shader_variant *list, unsigned **const_range
 
 static inline boolean
 nine_shader_variant_add(struct nine_shader_variant *list,
-                          uint64_t key, void *cso, unsigned *const_ranges)
+                        uint64_t key, void *cso,
+                        unsigned *const_ranges,
+                        unsigned const_used_size)
 {
     while (list->next) {
         assert(list->key != key);
@@ -149,6 +156,7 @@ nine_shader_variant_add(struct nine_shader_variant *list,
     list->next->key = key;
     list->next->cso = cso;
     list->next->const_ranges = const_ranges;
+    list->next->const_used_size = const_used_size;
     return TRUE;
 }
 
diff --git a/src/gallium/state_trackers/nine/nine_state.c b/src/gallium/state_trackers/nine/nine_state.c
index 6440cf46c93..7de5fc9a4c6 100644
--- a/src/gallium/state_trackers/nine/nine_state.c
+++ b/src/gallium/state_trackers/nine/nine_state.c
@@ -437,7 +437,7 @@ prepare_vs_constants_userbuf(struct NineDevice9 *device)
     struct pipe_constant_buffer cb;
     cb.buffer = NULL;
     cb.buffer_offset = 0;
-    cb.buffer_size = context->vs->const_used_size;
+    cb.buffer_size = context->cso_shader.vs_const_used_size;
     cb.user_buffer = context->vs_const_f;
 
     if (context->swvp) {
@@ -518,7 +518,7 @@ prepare_ps_constants_userbuf(struct NineDevice9 *device)
     struct pipe_constant_buffer cb;
     cb.buffer = NULL;
     cb.buffer_offset = 0;
-    cb.buffer_size = context->ps->const_used_size;
+    cb.buffer_size = context->cso_shader.ps_const_used_size;
     cb.user_buffer = context->ps_const_f;
 
     if (context->changed.ps_const_i) {
@@ -535,7 +535,7 @@ prepare_ps_constants_userbuf(struct NineDevice9 *device)
 
     /* Upload special constants needed to implement PS1.x instructions like TEXBEM,TEXBEML and BEM */
     if (context->ps->bumpenvmat_needed) {
-        memcpy(context->ps_lconstf_temp, cb.user_buffer, cb.buffer_size);
+        memcpy(context->ps_lconstf_temp, cb.user_buffer, 8 * sizeof(float[4]));
         memcpy(&context->ps_lconstf_temp[4 * 8], &device->context.bumpmap_vars, sizeof(device->context.bumpmap_vars));
 
         cb.user_buffer = context->ps_lconstf_temp;
@@ -545,7 +545,7 @@ prepare_ps_constants_userbuf(struct NineDevice9 *device)
         context->rs[D3DRS_FOGENABLE]) {
         float *dst = &context->ps_lconstf_temp[4 * 32];
         if (cb.user_buffer != context->ps_lconstf_temp) {
-            memcpy(context->ps_lconstf_temp, cb.user_buffer, cb.buffer_size);
+            memcpy(context->ps_lconstf_temp, cb.user_buffer, 32 * sizeof(float[4]));
             cb.user_buffer = context->ps_lconstf_temp;
         }
 
@@ -556,7 +556,6 @@ prepare_ps_constants_userbuf(struct NineDevice9 *device)
         } else if (context->rs[D3DRS_FOGTABLEMODE] != D3DFOG_NONE) {
             dst[4] = asfloat(context->rs[D3DRS_FOGDENSITY]);
         }
-        cb.buffer_size = 4 * 4 * 34;
     }
 
     if (!cb.buffer_size)
@@ -603,7 +602,9 @@ prepare_vs(struct NineDevice9 *device, uint8_t shader_changed)
 
     /* likely because we dislike FF */
     if (likely(context->programmable_vs)) {
-        context->cso_shader.vs = NineVertexShader9_GetVariant(vs, &context->cso_shader.vs_const_ranges);
+        context->cso_shader.vs = NineVertexShader9_GetVariant(vs,
+                                                              &context->cso_shader.vs_const_ranges,
+                                                              &context->cso_shader.vs_const_used_size);
     } else {
         vs = device->ff.vs;
         context->cso_shader.vs = vs->ff_cso;
@@ -637,7 +638,9 @@ prepare_ps(struct NineDevice9 *device, uint8_t shader_changed)
         return 0;
 
     if (likely(ps)) {
-        context->cso_shader.ps = NinePixelShader9_GetVariant(ps, &context->cso_shader.ps_const_ranges);
+        context->cso_shader.ps = NinePixelShader9_GetVariant(ps,
+                                                             &context->cso_shader.ps_const_ranges,
+                                                             &context->cso_shader.ps_const_used_size);
     } else {
         ps = device->ff.ps;
         context->cso_shader.ps = ps->ff_cso;
diff --git a/src/gallium/state_trackers/nine/nine_state.h b/src/gallium/state_trackers/nine/nine_state.h
index 37238b85936..d8fa2f80c4d 100644
--- a/src/gallium/state_trackers/nine/nine_state.h
+++ b/src/gallium/state_trackers/nine/nine_state.h
@@ -240,8 +240,10 @@ struct nine_context {
     struct {
         void *vs;
         unsigned *vs_const_ranges;
+        unsigned vs_const_used_size;
         void *ps;
         unsigned *ps_const_ranges;
+        unsigned ps_const_used_size;
     } cso_shader;
 
     struct pipe_context *pipe;
diff --git a/src/gallium/state_trackers/nine/pixelshader9.c b/src/gallium/state_trackers/nine/pixelshader9.c
index 42ad2c76a7f..4b85c738f79 100644
--- a/src/gallium/state_trackers/nine/pixelshader9.c
+++ b/src/gallium/state_trackers/nine/pixelshader9.c
@@ -80,13 +80,14 @@ NinePixelShader9_ctor( struct NinePixelShader9 *This,
 
     This->variant.cso = info.cso;
     This->variant.const_ranges = info.const_ranges;
+    This->variant.const_used_size = info.const_used_size;
     This->last_cso = info.cso;
     This->last_const_ranges = info.const_ranges;
+    This->last_const_used_size = info.const_used_size;
     This->last_key = 0;
 
     This->sampler_mask = info.sampler_mask;
     This->rt_mask = info.rt_mask;
-    This->const_used_size = info.const_used_size;
     This->bumpenvmat_needed = info.bumpenvmat_needed;
 
     memcpy(This->int_slots_used, info.int_slots_used, sizeof(This->int_slots_used));
@@ -159,7 +160,9 @@ NinePixelShader9_GetFunction( struct NinePixelShader9 *This,
 }
 
 void *
-NinePixelShader9_GetVariant( struct NinePixelShader9 *This, unsigned **const_ranges )
+NinePixelShader9_GetVariant( struct NinePixelShader9 *This,
+                             unsigned **const_ranges,
+                             unsigned *const_used_size )
 {
     /* GetVariant is called from nine_context, thus we can
      * get pipe directly */
@@ -170,10 +173,11 @@ NinePixelShader9_GetVariant( struct NinePixelShader9 *This, unsigned **const_ran
     key = This->next_key;
     if (key == This->last_key) {
         *const_ranges = This->last_const_ranges;
+        *const_used_size = This->last_const_used_size;
         return This->last_cso;
     }
 
-    cso = nine_shader_variant_get(&This->variant, const_ranges, key);
+    cso = nine_shader_variant_get(&This->variant, const_ranges, const_used_size, key);
     if (!cso) {
         struct NineDevice9 *device = This->base.device;
         struct nine_shader_info info;
@@ -210,14 +214,17 @@ NinePixelShader9_GetVariant( struct NinePixelShader9 *This, unsigned **const_ran
         hr = nine_translate_shader(This->base.device, &info, pipe);
         if (FAILED(hr))
             return NULL;
-        nine_shader_variant_add(&This->variant, key, info.cso, info.const_ranges);
+        nine_shader_variant_add(&This->variant, key, info.cso,
+                                info.const_ranges, info.const_used_size);
         cso = info.cso;
         *const_ranges = info.const_ranges;
+        *const_used_size = info.const_used_size;
     }
 
     This->last_key = key;
     This->last_cso = cso;
     This->last_const_ranges = *const_ranges;
+    This->last_const_used_size = *const_used_size;
 
     return cso;
 }
diff --git a/src/gallium/state_trackers/nine/pixelshader9.h b/src/gallium/state_trackers/nine/pixelshader9.h
index 62b5abb5962..6bac90be7a9 100644
--- a/src/gallium/state_trackers/nine/pixelshader9.h
+++ b/src/gallium/state_trackers/nine/pixelshader9.h
@@ -43,8 +43,6 @@ struct NinePixelShader9
         uint8_t version; /* (major << 4) | minor */
     } byte_code;
 
-    unsigned const_used_size; /* in bytes */
-
     uint8_t bumpenvmat_needed;
     uint16_t sampler_mask;
     uint8_t rt_mask;
@@ -63,6 +61,7 @@ struct NinePixelShader9
     uint64_t last_key;
     void *last_cso;
     unsigned *last_const_ranges;
+    unsigned last_const_used_size; /* in bytes */
 
     uint64_t next_key;
 };
@@ -132,7 +131,9 @@ NinePixelShader9_UpdateKey( struct NinePixelShader9 *ps,
 }
 
 void *
-NinePixelShader9_GetVariant( struct NinePixelShader9 *ps, unsigned **const_ranges );
+NinePixelShader9_GetVariant( struct NinePixelShader9 *ps,
+                             unsigned **const_ranges,
+                             unsigned *const_used_size );
 
 /*** public ***/
 
diff --git a/src/gallium/state_trackers/nine/vertexshader9.c b/src/gallium/state_trackers/nine/vertexshader9.c
index d14b9702a18..04c50ae619f 100644
--- a/src/gallium/state_trackers/nine/vertexshader9.c
+++ b/src/gallium/state_trackers/nine/vertexshader9.c
@@ -95,11 +95,12 @@ NineVertexShader9_ctor( struct NineVertexShader9 *This,
 
     This->variant.cso = info.cso;
     This->variant.const_ranges = info.const_ranges;
+    This->variant.const_used_size = info.const_used_size;
     This->last_cso = info.cso;
     This->last_const_ranges = info.const_ranges;
+    This->last_const_used_size = info.const_used_size;
     This->last_key = (uint32_t) (info.swvp_on << 9);
 
-    This->const_used_size = info.const_used_size;
     This->lconstf = info.lconstf;
     This->sampler_mask = info.sampler_mask;
     This->position_t = info.position_t;
@@ -185,7 +186,9 @@ NineVertexShader9_GetFunction( struct NineVertexShader9 *This,
 }
 
 void *
-NineVertexShader9_GetVariant( struct NineVertexShader9 *This, unsigned **const_ranges )
+NineVertexShader9_GetVariant( struct NineVertexShader9 *This,
+                              unsigned **const_ranges,
+                              unsigned *const_used_size )
 {
     /* GetVariant is called from nine_context, thus we can
      * get pipe directly */
@@ -196,10 +199,11 @@ NineVertexShader9_GetVariant( struct NineVertexShader9 *This, unsigned **const_r
     key = This->next_key;
     if (key == This->last_key) {
         *const_ranges = This->last_const_ranges;
+        *const_used_size = This->last_const_used_size;
         return This->last_cso;
     }
 
-    cso = nine_shader_variant_get(&This->variant, const_ranges, key);
+    cso = nine_shader_variant_get(&This->variant, const_ranges, const_used_size, key);
     if (!cso) {
         struct NineDevice9 *device = This->base.device;
         struct nine_shader_info info;
@@ -223,14 +227,17 @@ NineVertexShader9_GetVariant( struct NineVertexShader9 *This, unsigned **const_r
         hr = nine_translate_shader(This->base.device, &info, pipe);
         if (FAILED(hr))
             return NULL;
-        nine_shader_variant_add(&This->variant, key, info.cso, info.const_ranges);
+        nine_shader_variant_add(&This->variant, key, info.cso,
+                                info.const_ranges, info.const_used_size);
         cso = info.cso;
         *const_ranges = info.const_ranges;
+        *const_used_size = info.const_used_size;
     }
 
     This->last_key = key;
     This->last_cso = cso;
     This->last_const_ranges = *const_ranges;
+    This->last_const_used_size = *const_used_size;
 
     return cso;
 }
diff --git a/src/gallium/state_trackers/nine/vertexshader9.h b/src/gallium/state_trackers/nine/vertexshader9.h
index 0b139c5dad1..cbbd34979c9 100644
--- a/src/gallium/state_trackers/nine/vertexshader9.h
+++ b/src/gallium/state_trackers/nine/vertexshader9.h
@@ -55,8 +55,6 @@ struct NineVertexShader9
     boolean point_size; /* if true, set rasterizer.point_size_per_vertex to 1 */
     boolean swvp_only;
 
-    unsigned const_used_size; /* in bytes */
-
     struct nine_lconstf lconstf;
 
     boolean int_slots_used[NINE_MAX_CONST_I];
@@ -73,6 +71,7 @@ struct NineVertexShader9
     uint64_t last_key;
     void *last_cso;
     unsigned *last_const_ranges;
+    unsigned last_const_used_size; /* in bytes */
 
     uint64_t next_key;
 
@@ -124,7 +123,9 @@ NineVertexShader9_UpdateKey( struct NineVertexShader9 *vs,
 }
 
 void *
-NineVertexShader9_GetVariant( struct NineVertexShader9 *vs, unsigned **const_ranges );
+NineVertexShader9_GetVariant( struct NineVertexShader9 *vs,
+                              unsigned **const_ranges,
+                              unsigned *const_used_size );
 
 void *
 NineVertexShader9_GetVariantProcessVertices( struct NineVertexShader9 *vs,