#include "gallivm/lp_bld_gather.h"
#include "tgsi/tgsi_parse.h"
#include "util/u_double_list.h"
+#include "util/u_memory.h"
#include "r600.h"
#include "r600_asm.h"
#include "r600_opcodes.h"
#include "r600_shader.h"
+#include "r600_pipe.h"
#include "radeon_llvm.h"
#include "radeon_llvm_emit.h"
#include <stdio.h>
+#if defined R600_USE_LLVM || defined HAVE_OPENCL
+
static LLVMValueRef llvm_fetch_const(
struct lp_build_tgsi_context * bld_base,
const struct tgsi_full_src_register *reg,
enum tgsi_opcode_type type,
unsigned swizzle)
{
- LLVMValueRef cval = lp_build_intrinsic_unary(bld_base->base.gallivm->builder,
+ LLVMValueRef idx = lp_build_const_int32(bld_base->base.gallivm,
+ radeon_llvm_reg_index_soa(reg->Register.Index, swizzle));
+ LLVMValueRef cval = build_intrinsic(bld_base->base.gallivm->builder,
"llvm.AMDGPU.load.const", bld_base->base.elem_type,
- lp_build_const_int32(bld_base->base.gallivm,
- radeon_llvm_reg_index_soa(reg->Register.Index, swizzle)));
+ &idx, 1, LLVMReadNoneAttribute);
return bitcast(bld_base, type, cval);
}
LLVMValueRef reg = lp_build_const_int32(
ctx->soa.bld_base.base.gallivm, chan);
- ctx->system_values[index] = lp_build_intrinsic_unary(
+ ctx->system_values[index] = build_intrinsic(
ctx->soa.bld_base.base.gallivm->builder,
"llvm.R600.load.input",
- ctx->soa.bld_base.base.elem_type, reg);
+ ctx->soa.bld_base.base.elem_type, ®, 1,
+ LLVMReadNoneAttribute);
}
static LLVMValueRef llvm_fetch_system_value(
return bitcast(bld_base, type, cval);
}
+static LLVMValueRef
+llvm_load_input_helper(
+ struct radeon_llvm_context * ctx,
+ const char *intrinsic, unsigned idx)
+{
+ LLVMValueRef reg = lp_build_const_int32(
+ ctx->soa.bld_base.base.gallivm,
+ idx);
+ return build_intrinsic(
+ ctx->soa.bld_base.base.gallivm->builder,
+ intrinsic,
+ ctx->soa.bld_base.base.elem_type, ®, 1,
+ LLVMReadNoneAttribute);
+}
+
+static LLVMValueRef
+llvm_face_select_helper(
+ struct radeon_llvm_context * ctx,
+ const char *intrinsic, unsigned face_register,
+ unsigned frontcolor_register, unsigned backcolor_regiser)
+{
+
+ LLVMValueRef backcolor = llvm_load_input_helper(
+ ctx,
+ intrinsic,
+ backcolor_regiser);
+ LLVMValueRef front_color = llvm_load_input_helper(
+ ctx,
+ intrinsic,
+ frontcolor_register);
+ LLVMValueRef face = llvm_load_input_helper(
+ ctx,
+ "llvm.R600.load.input",
+ face_register);
+ LLVMValueRef is_face_positive = LLVMBuildFCmp(
+ ctx->soa.bld_base.base.gallivm->builder,
+ LLVMRealUGT, face,
+ lp_build_const_float(ctx->soa.bld_base.base.gallivm, 0.0f),
+ "");
+ return LLVMBuildSelect(
+ ctx->soa.bld_base.base.gallivm->builder,
+ is_face_positive,
+ front_color,
+ backcolor,
+ "");
+}
+
static void llvm_load_input(
struct radeon_llvm_context * ctx,
unsigned input_index,
{
unsigned chan;
+ const char *intrinsics = "llvm.R600.load.input";
+ unsigned offset = 4 * ctx->reserved_reg_count;
+
+ if (ctx->type == TGSI_PROCESSOR_FRAGMENT && ctx->chip_class >= EVERGREEN) {
+ switch (decl->Interp.Interpolate) {
+ case TGSI_INTERPOLATE_COLOR:
+ case TGSI_INTERPOLATE_PERSPECTIVE:
+ offset = 0;
+ intrinsics = "llvm.R600.load.input.perspective";
+ break;
+ case TGSI_INTERPOLATE_LINEAR:
+ offset = 0;
+ intrinsics = "llvm.R600.load.input.linear";
+ break;
+ case TGSI_INTERPOLATE_CONSTANT:
+ offset = 0;
+ intrinsics = "llvm.R600.load.input.constant";
+ break;
+ default:
+ assert(0 && "Unknow Interpolate mode");
+ }
+ }
+
for (chan = 0; chan < 4; chan++) {
unsigned soa_index = radeon_llvm_reg_index_soa(input_index,
chan);
- /* The * 4 is assuming that we are in soa mode. */
- LLVMValueRef reg = lp_build_const_int32(
- ctx->soa.bld_base.base.gallivm,
- soa_index + (ctx->reserved_reg_count * 4));
- ctx->inputs[soa_index] = lp_build_intrinsic_unary(
- ctx->soa.bld_base.base.gallivm->builder,
+ switch (decl->Semantic.Name) {
+ case TGSI_SEMANTIC_FACE:
+ ctx->inputs[soa_index] = llvm_load_input_helper(ctx,
+ "llvm.R600.load.input",
+ 4 * ctx->face_input);
+ break;
+ case TGSI_SEMANTIC_POSITION:
+ if (ctx->type != TGSI_PROCESSOR_FRAGMENT || chan != 3) {
+ ctx->inputs[soa_index] = llvm_load_input_helper(ctx,
+ "llvm.R600.load.input",
+ soa_index + (ctx->reserved_reg_count * 4));
+ } else {
+ LLVMValueRef w_coord = llvm_load_input_helper(ctx,
"llvm.R600.load.input",
- ctx->soa.bld_base.base.elem_type, reg);
+ soa_index + (ctx->reserved_reg_count * 4));
+ ctx->inputs[soa_index] = LLVMBuildFDiv(ctx->gallivm.builder,
+ lp_build_const_float(&(ctx->gallivm), 1.0f), w_coord, "");
+ }
+ break;
+ case TGSI_SEMANTIC_COLOR:
+ if (ctx->two_side) {
+ unsigned front_location, back_location;
+ unsigned back_reg = ctx->r600_inputs[input_index]
+ .potential_back_facing_reg;
+ if (ctx->chip_class >= EVERGREEN) {
+ front_location = 4 * ctx->r600_inputs[input_index].lds_pos + chan;
+ back_location = 4 * ctx->r600_inputs[back_reg].lds_pos + chan;
+ } else {
+ front_location = soa_index + 4 * ctx->reserved_reg_count;
+ back_location = radeon_llvm_reg_index_soa(
+ ctx->r600_inputs[back_reg].gpr,
+ chan);
+ }
+ ctx->inputs[soa_index] = llvm_face_select_helper(ctx,
+ intrinsics,
+ 4 * ctx->face_input, front_location, back_location);
+ break;
+ }
+ default:
+ {
+ unsigned location;
+ if (ctx->chip_class >= EVERGREEN) {
+ location = 4 * ctx->r600_inputs[input_index].lds_pos + chan;
+ } else {
+ location = soa_index + 4 * ctx->reserved_reg_count;
+ }
+ /* The * 4 is assuming that we are in soa mode. */
+ ctx->inputs[soa_index] = llvm_load_input_helper(ctx,
+ intrinsics, location);
+
+ break;
+ }
+ }
}
}
struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
struct lp_build_context * base = &bld_base->base;
unsigned i;
+
+ unsigned color_count = 0;
+ boolean has_color = false;
/* Add the necessary export instructions */
for (i = 0; i < ctx->output_reg_count; i++) {
LLVMValueRef output;
unsigned adjusted_reg_idx = i +
ctx->reserved_reg_count;
- LLVMValueRef reg_index = lp_build_const_int32(
- base->gallivm,
- radeon_llvm_reg_index_soa(adjusted_reg_idx, chan));
output = LLVMBuildLoad(base->gallivm->builder,
ctx->soa.outputs[i][chan], "");
- lp_build_intrinsic_binary(
- base->gallivm->builder,
- "llvm.AMDGPU.store.output",
- LLVMVoidTypeInContext(base->gallivm->context),
- output, reg_index);
+ if (ctx->type == TGSI_PROCESSOR_VERTEX) {
+ LLVMValueRef reg_index = lp_build_const_int32(
+ base->gallivm,
+ radeon_llvm_reg_index_soa(adjusted_reg_idx, chan));
+ lp_build_intrinsic_binary(
+ base->gallivm->builder,
+ "llvm.AMDGPU.store.output",
+ LLVMVoidTypeInContext(base->gallivm->context),
+ output, reg_index);
+ } else if (ctx->type == TGSI_PROCESSOR_FRAGMENT) {
+ switch (ctx->r600_outputs[i].name) {
+ case TGSI_SEMANTIC_COLOR:
+ has_color = true;
+ if ( color_count/4 < ctx->color_buffer_count) {
+ if (ctx->fs_color_all) {
+ for (unsigned j = 0; j < ctx->color_buffer_count; j++) {
+ LLVMValueRef reg_index = lp_build_const_int32(
+ base->gallivm,
+ (j * 4) + chan);
+ lp_build_intrinsic_binary(
+ base->gallivm->builder,
+ "llvm.R600.store.pixel.color",
+ LLVMVoidTypeInContext(base->gallivm->context),
+ output, reg_index);
+ }
+ } else {
+ LLVMValueRef reg_index = lp_build_const_int32(
+ base->gallivm,
+ (color_count++/4) * 4 + chan);
+ lp_build_intrinsic_binary(
+ base->gallivm->builder,
+ "llvm.R600.store.pixel.color",
+ LLVMVoidTypeInContext(base->gallivm->context),
+ output, reg_index);
+ }
+ }
+ break;
+ case TGSI_SEMANTIC_POSITION:
+ if (chan != 2)
+ continue;
+ lp_build_intrinsic_unary(
+ base->gallivm->builder,
+ "llvm.R600.store.pixel.depth",
+ LLVMVoidTypeInContext(base->gallivm->context),
+ output);
+ break;
+ case TGSI_SEMANTIC_STENCIL:
+ if (chan != 1)
+ continue;
+ lp_build_intrinsic_unary(
+ base->gallivm->builder,
+ "llvm.R600.store.pixel.stencil",
+ LLVMVoidTypeInContext(base->gallivm->context),
+ output);
+ break;
+ }
+ }
}
}
+
+ if (!has_color && ctx->type == TGSI_PROCESSOR_FRAGMENT)
+ lp_build_intrinsic(base->gallivm->builder, "llvm.R600.store.pixel.dummy", LLVMVoidTypeInContext(base->gallivm->context), 0, 0);
}
static void llvm_emit_tex(
struct lp_build_emit_data * emit_data)
{
struct gallivm_state * gallivm = bld_base->base.gallivm;
- LLVMValueRef args[3];
+ LLVMValueRef args[6];
+ unsigned c, sampler_src;
+
+ assert(emit_data->arg_count + 2 <= Elements(args));
+
+ for (c = 0; c < emit_data->arg_count; ++c)
+ args[c] = emit_data->args[c];
- args[0] = emit_data->args[0];
- args[1] = lp_build_const_int32(gallivm,
- emit_data->inst->Src[1].Register.Index);
- args[2] = lp_build_const_int32(gallivm,
+ sampler_src = emit_data->inst->Instruction.NumSrcRegs-1;
+
+ args[c++] = lp_build_const_int32(gallivm,
+ emit_data->inst->Src[sampler_src].Register.Index + R600_MAX_CONST_BUFFERS);
+ args[c++] = lp_build_const_int32(gallivm,
+ emit_data->inst->Src[sampler_src].Register.Index);
+ args[c++] = lp_build_const_int32(gallivm,
emit_data->inst->Texture.Texture);
- emit_data->output[0] = lp_build_intrinsic(gallivm->builder,
+
+ emit_data->output[0] = build_intrinsic(gallivm->builder,
action->intr_name,
- emit_data->dst_type, args, 3);
+ emit_data->dst_type, args, c, LLVMReadNoneAttribute);
+}
+
+static void emit_cndlt(
+ const struct lp_build_tgsi_action * action,
+ struct lp_build_tgsi_context * bld_base,
+ struct lp_build_emit_data * emit_data)
+{
+ LLVMBuilderRef builder = bld_base->base.gallivm->builder;
+ LLVMValueRef float_zero = lp_build_const_float(
+ bld_base->base.gallivm, 0.0f);
+ LLVMValueRef cmp = LLVMBuildFCmp(
+ builder, LLVMRealULT, emit_data->args[0], float_zero, "");
+ emit_data->output[emit_data->chan] = LLVMBuildSelect(builder,
+ cmp, emit_data->args[1], emit_data->args[2], "");
}
static void dp_fetch_args(
static struct lp_build_tgsi_action dot_action = {
.fetch_args = dp_fetch_args,
- .emit = lp_build_tgsi_intrinsic,
+ .emit = build_tgsi_intrinsic_nomem,
.intr_name = "llvm.AMDGPU.dp4"
};
bld_base->op_actions[TGSI_OPCODE_DDX].emit = llvm_emit_tex;
bld_base->op_actions[TGSI_OPCODE_DDY].emit = llvm_emit_tex;
bld_base->op_actions[TGSI_OPCODE_TEX].emit = llvm_emit_tex;
+ bld_base->op_actions[TGSI_OPCODE_TEX2].emit = llvm_emit_tex;
bld_base->op_actions[TGSI_OPCODE_TXB].emit = llvm_emit_tex;
+ bld_base->op_actions[TGSI_OPCODE_TXB2].emit = llvm_emit_tex;
bld_base->op_actions[TGSI_OPCODE_TXD].emit = llvm_emit_tex;
bld_base->op_actions[TGSI_OPCODE_TXL].emit = llvm_emit_tex;
+ bld_base->op_actions[TGSI_OPCODE_TXL2].emit = llvm_emit_tex;
bld_base->op_actions[TGSI_OPCODE_TXF].emit = llvm_emit_tex;
bld_base->op_actions[TGSI_OPCODE_TXQ].emit = llvm_emit_tex;
bld_base->op_actions[TGSI_OPCODE_TXP].emit = llvm_emit_tex;
+ bld_base->op_actions[TGSI_OPCODE_CMP].emit = emit_cndlt;
lp_build_tgsi_llvm(bld_base, tokens);
case CHIP_RV635:
case CHIP_RS780:
case CHIP_RS880:
+ gpu_family = "r600";
+ break;
case CHIP_RV710:
gpu_family = "rv710";
break;
return radeon_llvm_compile(mod, inst_bytes, inst_byte_count,
gpu_family, dump);
}
+
+#endif