ac: replace old image intrinsics with new ones
authorMarek Olšák <marek.olsak@amd.com>
Thu, 23 Feb 2017 22:37:59 +0000 (23:37 +0100)
committerMarek Olšák <marek.olsak@amd.com>
Fri, 3 Mar 2017 14:29:30 +0000 (15:29 +0100)
Reviewed-by: Dave Airlie <airlied@redhat.com>
src/amd/common/ac_llvm_build.c

index 3a1ef938e9928d881ea3ac849ec46541f4ce058e..187c2cb05e2904b43703fcaa16bce715fdde0b7e 100644 (file)
@@ -114,6 +114,20 @@ ac_emit_llvm_intrinsic(struct ac_llvm_context *ctx, const char *name,
        return call;
 }
 
+static LLVMValueRef bitcast_to_float(struct ac_llvm_context *ctx,
+                                    LLVMValueRef value)
+{
+       LLVMTypeRef type = LLVMTypeOf(value);
+       LLVMTypeRef new_type;
+
+       if (LLVMGetTypeKind(type) == LLVMVectorTypeKind)
+               new_type = LLVMVectorType(ctx->f32, LLVMGetVectorSize(type));
+       else
+               new_type = ctx->f32;
+
+       return LLVMBuildBitCast(ctx->builder, value, new_type, "");
+}
+
 /**
  * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
  * intrinsic names).
@@ -942,6 +956,72 @@ LLVMValueRef ac_emit_image_opcode(struct ac_llvm_context *ctx,
        const char *name;
        char intr_name[128], type[64];
 
+       if (HAVE_LLVM >= 0x0400) {
+               bool sample = a->opcode == ac_image_sample ||
+                             a->opcode == ac_image_gather4 ||
+                             a->opcode == ac_image_get_lod;
+
+               if (sample)
+                       args[num_args++] = bitcast_to_float(ctx, a->addr);
+               else
+                       args[num_args++] = a->addr;
+
+               args[num_args++] = a->resource;
+               if (sample)
+                       args[num_args++] = a->sampler;
+               args[num_args++] = LLVMConstInt(ctx->i32, a->dmask, 0);
+               if (sample)
+                       args[num_args++] = LLVMConstInt(ctx->i1, a->unorm, 0);
+               args[num_args++] = LLVMConstInt(ctx->i1, 0, 0); /* glc */
+               args[num_args++] = LLVMConstInt(ctx->i1, 0, 0); /* slc */
+               args[num_args++] = LLVMConstInt(ctx->i1, 0, 0); /* lwe */
+               args[num_args++] = LLVMConstInt(ctx->i1, a->da, 0);
+
+               switch (a->opcode) {
+               case ac_image_sample:
+                       name = "llvm.amdgcn.image.sample";
+                       break;
+               case ac_image_gather4:
+                       name = "llvm.amdgcn.image.gather4";
+                       break;
+               case ac_image_load:
+                       name = "llvm.amdgcn.image.load";
+                       break;
+               case ac_image_load_mip:
+                       name = "llvm.amdgcn.image.load.mip";
+                       break;
+               case ac_image_get_lod:
+                       name = "llvm.amdgcn.image.getlod";
+                       break;
+               case ac_image_get_resinfo:
+                       name = "llvm.amdgcn.image.getresinfo";
+                       break;
+               }
+
+               ac_build_type_name_for_intr(LLVMTypeOf(args[0]), type,
+                                           sizeof(type));
+
+               snprintf(intr_name, sizeof(intr_name), "%s%s%s%s.v4f32.%s.v8i32",
+                       name,
+                       a->compare ? ".c" : "",
+                       a->bias ? ".b" :
+                       a->lod ? ".l" :
+                       a->deriv ? ".d" :
+                       a->level_zero ? ".lz" : "",
+                       a->offset ? ".o" : "",
+                       type);
+
+               LLVMValueRef result =
+                       ac_emit_llvm_intrinsic(ctx, intr_name,
+                                              ctx->v4f32, args, num_args,
+                                              AC_FUNC_ATTR_READNONE);
+               if (!sample) {
+                       result = LLVMBuildBitCast(ctx->builder, result,
+                                                 ctx->v4i32, "");
+               }
+               return result;
+       }
+
        args[num_args++] = a->addr;
        args[num_args++] = a->resource;