radeonsi: switch descriptors to i32 vectors
authorMarek Olšák <marek.olsak@amd.com>
Mon, 7 Jul 2014 20:49:15 +0000 (22:49 +0200)
committerMarek Olšák <marek.olsak@amd.com>
Fri, 11 Jul 2014 17:36:29 +0000 (19:36 +0200)
This is a follow-up to the commit which adds texture fetches with offsets.

Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
src/gallium/drivers/radeonsi/si_shader.c

index 4c0309b1390602a5a55d65ea4501932c174a009f..5ac38cd9996b65fdc2f1880e04bca205afd17537 100644 (file)
@@ -1574,7 +1574,7 @@ static void tex_fetch_args(
                LLVMTypeRef i8 = LLVMInt8TypeInContext(gallivm->context);
                LLVMTypeRef v16i8 = LLVMVectorType(i8, 16);
 
-               /* Truncate v32i8 to v16i8. */
+               /* Bitcast and truncate v8i32 to v16i8. */
                LLVMValueRef res = si_shader_ctx->resources[sampler_index];
                res = LLVMBuildBitCast(gallivm->builder, res, v2i128, "");
                res = LLVMBuildExtractElement(gallivm->builder, res, bld_base->uint_bld.zero, "");
@@ -2305,12 +2305,18 @@ static void create_meta_data(struct si_shader_context *si_shader_ctx)
        si_shader_ctx->const_md = LLVMMDNodeInContext(gallivm->context, args, 3);
 }
 
+static LLVMTypeRef const_array(LLVMTypeRef elem_type, int num_elements)
+{
+       return LLVMPointerType(LLVMArrayType(elem_type, num_elements),
+                              CONST_ADDR_SPACE);
+}
+
 static void create_function(struct si_shader_context *si_shader_ctx)
 {
        struct lp_build_tgsi_context *bld_base = &si_shader_ctx->radeon_bld.soa.bld_base;
        struct gallivm_state *gallivm = bld_base->base.gallivm;
        struct si_pipe_shader *shader = si_shader_ctx->shader;
-       LLVMTypeRef params[SI_NUM_PARAMS], f32, i8, i32, v2i32, v3i32;
+       LLVMTypeRef params[SI_NUM_PARAMS], f32, i8, i32, v2i32, v3i32, v16i8, v4i32, v8i32;
        unsigned i, last_sgpr, num_params;
 
        i8 = LLVMInt8TypeInContext(gallivm->context);
@@ -2318,21 +2324,18 @@ static void create_function(struct si_shader_context *si_shader_ctx)
        f32 = LLVMFloatTypeInContext(gallivm->context);
        v2i32 = LLVMVectorType(i32, 2);
        v3i32 = LLVMVectorType(i32, 3);
+       v4i32 = LLVMVectorType(i32, 4);
+       v8i32 = LLVMVectorType(i32, 8);
+       v16i8 = LLVMVectorType(i8, 16);
 
-       params[SI_PARAM_CONST] = LLVMPointerType(
-               LLVMArrayType(LLVMVectorType(i8, 16), NUM_CONST_BUFFERS), CONST_ADDR_SPACE);
-       params[SI_PARAM_RW_BUFFERS] = params[SI_PARAM_CONST];
-
-       /* We assume at most 16 textures per program at the moment.
-        * This need probably need to be changed to support bindless textures */
-       params[SI_PARAM_SAMPLER] = LLVMPointerType(
-               LLVMArrayType(LLVMVectorType(i8, 16), NUM_SAMPLER_STATES), CONST_ADDR_SPACE);
-       params[SI_PARAM_RESOURCE] = LLVMPointerType(
-               LLVMArrayType(LLVMVectorType(i8, 32), NUM_SAMPLER_VIEWS), CONST_ADDR_SPACE);
+       params[SI_PARAM_CONST] = const_array(v16i8, NUM_CONST_BUFFERS);
+       params[SI_PARAM_RW_BUFFERS] = const_array(v16i8, 6); /* XXX hardcoded */
+       params[SI_PARAM_SAMPLER] = const_array(v4i32, NUM_SAMPLER_STATES);
+       params[SI_PARAM_RESOURCE] = const_array(v8i32, NUM_SAMPLER_VIEWS);
 
        switch (si_shader_ctx->type) {
        case TGSI_PROCESSOR_VERTEX:
-               params[SI_PARAM_VERTEX_BUFFER] = params[SI_PARAM_CONST];
+               params[SI_PARAM_VERTEX_BUFFER] = const_array(v16i8, 16); /* XXX hardcoded */
                params[SI_PARAM_START_INSTANCE] = i32;
                num_params = SI_PARAM_START_INSTANCE+1;
                if (shader->key.vs.as_es) {