gallium/util: s/uint/enum tgsi_semantic/ in simple shader code
[mesa.git] / src / gallium / auxiliary / draw / draw_llvm.c
index 19b75a5003bfd21c579e12f090d666372ad10c8f..69c108037509e61243d2663b77ea9da2a46da649 100644 (file)
@@ -352,9 +352,9 @@ create_jit_vertex_buffer_type(struct gallivm_state *gallivm,
    LLVMTypeRef elem_types[4];
    LLVMTypeRef vb_type;
 
-   elem_types[0] =
-   elem_types[1] = LLVMInt32TypeInContext(gallivm->context);
-   elem_types[2] =
+   elem_types[0] = LLVMInt16TypeInContext(gallivm->context);
+   elem_types[1] = LLVMInt8TypeInContext(gallivm->context);
+   elem_types[2] = LLVMInt32TypeInContext(gallivm->context);
    elem_types[3] = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
 
    vb_type = LLVMStructTypeInContext(gallivm->context, elem_types,
@@ -363,8 +363,12 @@ create_jit_vertex_buffer_type(struct gallivm_state *gallivm,
    (void) target; /* silence unused var warning for non-debug build */
    LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, stride,
                           target, vb_type, 0);
-   LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer_offset,
+   LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, is_user_buffer,
                           target, vb_type, 1);
+   LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer_offset,
+                          target, vb_type, 2);
+   LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer.resource,
+                          target, vb_type, 3);
 
    LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer, target, vb_type);
 
@@ -602,7 +606,7 @@ generate_vs(struct draw_llvm_variant *variant,
             const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
             const struct lp_bld_tgsi_system_values *system_values,
             LLVMValueRef context_ptr,
-            struct lp_build_sampler_soa *draw_sampler,
+            const struct lp_build_sampler_soa *draw_sampler,
             boolean clamp_vertex_color)
 {
    struct draw_llvm *llvm = variant->llvm;
@@ -712,39 +716,6 @@ fetch_instanced(struct gallivm_state *gallivm,
 }
 
 
-static void
-convert_to_soa(struct gallivm_state *gallivm,
-               LLVMValueRef src_aos[LP_MAX_VECTOR_WIDTH / 32],
-               LLVMValueRef dst_soa[TGSI_NUM_CHANNELS],
-               const struct lp_type soa_type)
-{
-   unsigned j, k;
-   struct lp_type aos_channel_type = soa_type;
-
-   LLVMValueRef aos_channels[TGSI_NUM_CHANNELS];
-   unsigned pixels_per_channel = soa_type.length / TGSI_NUM_CHANNELS;
-
-   debug_assert(TGSI_NUM_CHANNELS == 4);
-   debug_assert((soa_type.length % TGSI_NUM_CHANNELS) == 0);
-
-   aos_channel_type.length >>= 1;
-
-   for (j = 0; j < TGSI_NUM_CHANNELS; ++j) {
-      LLVMValueRef channel[LP_MAX_VECTOR_LENGTH] = { 0 };
-
-      assert(pixels_per_channel <= LP_MAX_VECTOR_LENGTH);
-
-      for (k = 0; k < pixels_per_channel; ++k) {
-         channel[k] = src_aos[j + TGSI_NUM_CHANNELS * k];
-      }
-
-      aos_channels[j] = lp_build_concat(gallivm, channel, aos_channel_type, pixels_per_channel);
-   }
-
-   lp_build_transpose_aos(gallivm, soa_type, aos_channels, dst_soa);
-}
-
-
 static void
 fetch_vector(struct gallivm_state *gallivm,
              const struct util_format_description *format_desc,
@@ -755,11 +726,10 @@ fetch_vector(struct gallivm_state *gallivm,
              LLVMValueRef *inputs,
              LLVMValueRef indices)
 {
-   LLVMValueRef zero = LLVMConstNull(LLVMInt32TypeInContext(gallivm->context));
    LLVMBuilderRef builder = gallivm->builder;
    struct lp_build_context blduivec;
+   struct lp_type fetch_type = vs_type;
    LLVMValueRef offset, valid_mask;
-   LLVMValueRef aos_fetch[LP_MAX_VECTOR_WIDTH / 32];
    unsigned i;
 
    lp_build_context_init(&blduivec, gallivm, lp_uint_type(vs_type));
@@ -783,28 +753,37 @@ fetch_vector(struct gallivm_state *gallivm,
    }
 
    /*
-    * Note: we probably really want to use SoA fetch, not AoS one (albeit
-    * for most formats it will amount to the same as this isn't very
-    * optimized). But looks dangerous since it assumes alignment.
+    * Unlike fetch_instanced, use SoA fetch instead of multiple AoS fetches.
+    * This should always produce better code.
     */
-   for (i = 0; i < vs_type.length; i++) {
-      LLVMValueRef offset1, elem;
-      elem = lp_build_const_int32(gallivm, i);
-      offset1 = LLVMBuildExtractElement(builder, offset, elem, "");
 
-      aos_fetch[i] = lp_build_fetch_rgba_aos(gallivm, format_desc,
-                                             lp_float32_vec4_type(),
-                                             FALSE, map_ptr, offset1,
-                                             zero, zero, NULL);
+   /* The type handling is annoying here... */
+   if (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB &&
+       format_desc->channel[0].pure_integer) {
+      if (format_desc->channel[0].type == UTIL_FORMAT_TYPE_SIGNED) {
+         fetch_type = lp_type_int_vec(vs_type.width, vs_type.width * vs_type.length);
+      }
+      else if (format_desc->channel[0].type == UTIL_FORMAT_TYPE_UNSIGNED) {
+         fetch_type = lp_type_uint_vec(vs_type.width, vs_type.width * vs_type.length);
+      }
+   }
+
+   lp_build_fetch_rgba_soa(gallivm, format_desc,
+                           fetch_type, FALSE, map_ptr, offset,
+                           blduivec.zero, blduivec.zero,
+                           NULL, inputs);
+
+   for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
+      inputs[i] = LLVMBuildBitCast(builder, inputs[i],
+                                   lp_build_vec_type(gallivm, vs_type), "");
    }
-   convert_to_soa(gallivm, aos_fetch, inputs, vs_type);
 
+   /* out-of-bound fetches return all zeros */
    for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
       inputs[i] = LLVMBuildBitCast(builder, inputs[i], blduivec.vec_type, "");
       inputs[i] = LLVMBuildAnd(builder, inputs[i], valid_mask, "");
       inputs[i] = LLVMBuildBitCast(builder, inputs[i],
                                    lp_build_vec_type(gallivm, vs_type), "");
-
    }
 }
 
@@ -1695,8 +1674,6 @@ draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant)
       ind_vec = LLVMBuildInsertElement(builder, ind_vec, index, index, "");
    }
 
-   fetch_max = lp_build_alloca(gallivm, int32_type, "fetch_max");
-
    have_elts = LLVMBuildICmp(builder, LLVMIntNE,
                              LLVMConstPointerNull(arg_types[10]), fetch_elts, "");
 
@@ -1726,6 +1703,8 @@ draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant)
          vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr, &vb_index, 1, "");
          vb_info = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, "");
          vb_stride[j] = draw_jit_vbuffer_stride(gallivm, vb_info);
+         vb_stride[j] = LLVMBuildZExt(gallivm->builder, vb_stride[j],
+                                      LLVMInt32TypeInContext(context), "");
          vb_buffer_offset = draw_jit_vbuffer_offset(gallivm, vb_info);
          map_ptr[j] = draw_jit_dvbuffer_map(gallivm, vbuffer_ptr);
          buffer_size = draw_jit_dvbuffer_size(gallivm, vbuffer_ptr);
@@ -2093,7 +2072,7 @@ draw_llvm_dump_variant_key(struct draw_llvm_variant_key *key)
 
 void
 draw_llvm_set_mapped_texture(struct draw_context *draw,
-                             unsigned shader_stage,
+                             enum pipe_shader_type shader_stage,
                              unsigned sview_idx,
                              uint32_t width, uint32_t height, uint32_t depth,
                              uint32_t first_level, uint32_t last_level,
@@ -2138,7 +2117,7 @@ draw_llvm_set_mapped_texture(struct draw_context *draw,
 
 void
 draw_llvm_set_sampler_state(struct draw_context *draw, 
-                            unsigned shader_type)
+                            enum pipe_shader_type shader_type)
 {
    unsigned i;
 
@@ -2177,6 +2156,11 @@ draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
 {
    struct draw_llvm *llvm = variant->llvm;
 
+   if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
+      debug_printf("Deleting VS variant: %u vs variants,\t%u total variants\n",
+                    variant->shader->variants_cached, llvm->nr_variants);
+   }
+
    gallivm_destroy(variant->gallivm);
 
    remove_from_list(&variant->list_item_local);
@@ -2439,6 +2423,11 @@ draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant *variant)
 {
    struct draw_llvm *llvm = variant->llvm;
 
+   if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
+      debug_printf("Deleting GS variant: %u gs variants,\t%u total variants\n",
+                    variant->shader->variants_cached, llvm->nr_gs_variants);
+   }
+
    gallivm_destroy(variant->gallivm);
 
    remove_from_list(&variant->list_item_local);