/*
* Copyright 2016 Advanced Micro Devices, Inc.
+ * All Rights Reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
*/
#include "si_shader_internal.h"
-#include "gallivm/lp_bld_const.h"
-#include "gallivm/lp_bld_intr.h"
-#include "gallivm/lp_bld_gather.h"
-#include "tgsi/tgsi_parse.h"
+#include "si_pipe.h"
+#include "ac_llvm_util.h"
-static void kill_if_fetch_args(struct lp_build_tgsi_context *bld_base,
- struct lp_build_emit_data *emit_data)
+void si_llvm_emit_kill(struct ac_shader_abi *abi, LLVMValueRef visible)
{
- const struct tgsi_full_instruction *inst = emit_data->inst;
- struct gallivm_state *gallivm = bld_base->base.gallivm;
- LLVMBuilderRef builder = gallivm->builder;
- unsigned i;
- LLVMValueRef conds[TGSI_NUM_CHANNELS];
+ struct si_shader_context *ctx = si_shader_context_from_abi(abi);
+ LLVMBuilderRef builder = ctx->ac.builder;
- for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
- LLVMValueRef value = lp_build_emit_fetch(bld_base, inst, 0, i);
- conds[i] = LLVMBuildFCmp(builder, LLVMRealOLT, value,
- bld_base->base.zero, "");
- }
+ if (ctx->shader->selector->force_correct_derivs_after_kill) {
+ /* Kill immediately while maintaining WQM. */
+ ac_build_kill_if_false(&ctx->ac,
+ ac_build_wqm_vote(&ctx->ac, visible));
- /* Or the conditions together */
- for (i = TGSI_NUM_CHANNELS - 1; i > 0; i--) {
- conds[i - 1] = LLVMBuildOr(builder, conds[i], conds[i - 1], "");
+ LLVMValueRef mask = LLVMBuildLoad(builder, ctx->postponed_kill, "");
+ mask = LLVMBuildAnd(builder, mask, visible, "");
+ LLVMBuildStore(builder, mask, ctx->postponed_kill);
+ return;
}
- emit_data->dst_type = LLVMVoidTypeInContext(gallivm->context);
- emit_data->arg_count = 1;
- emit_data->args[0] = LLVMBuildSelect(builder, conds[0],
- lp_build_const_float(gallivm, -1.0f),
- bld_base->base.zero, "");
+ ac_build_kill_if_false(&ctx->ac, visible);
}
static void kil_emit(const struct lp_build_tgsi_action *action,
struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
- unsigned i;
- for (i = 0; i < emit_data->arg_count; i++) {
- emit_data->output[i] = lp_build_intrinsic_unary(
- bld_base->base.gallivm->builder,
- action->intr_name,
- emit_data->dst_type, emit_data->args[i]);
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ LLVMValueRef visible;
+
+ if (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_KILL_IF) {
+ const struct tgsi_full_instruction *inst = emit_data->inst;
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ LLVMBuilderRef builder = ctx->ac.builder;
+ unsigned i;
+ LLVMValueRef conds[TGSI_NUM_CHANNELS];
+
+ for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
+ LLVMValueRef value = lp_build_emit_fetch(bld_base, inst, 0, i);
+ /* UGE because NaN shouldn't get killed */
+ conds[i] = LLVMBuildFCmp(builder, LLVMRealUGE, value,
+ ctx->ac.f32_0, "");
+ }
+
+ /* And the conditions together */
+ for (i = TGSI_NUM_CHANNELS - 1; i > 0; i--) {
+ conds[i - 1] = LLVMBuildAnd(builder, conds[i], conds[i - 1], "");
+ }
+ visible = conds[0];
+ } else {
+ assert(emit_data->inst->Instruction.Opcode == TGSI_OPCODE_KILL);
+ visible = ctx->i1false;
}
+
+ si_llvm_emit_kill(&ctx->abi, visible);
}
static void emit_icmp(const struct lp_build_tgsi_action *action,
struct lp_build_emit_data *emit_data)
{
unsigned pred;
- LLVMBuilderRef builder = bld_base->base.gallivm->builder;
- LLVMContextRef context = bld_base->base.gallivm->context;
+ struct si_shader_context *ctx = si_shader_context(bld_base);
switch (emit_data->inst->Instruction.Opcode) {
case TGSI_OPCODE_USEQ:
break;
}
- LLVMValueRef v = LLVMBuildICmp(builder, pred,
+ LLVMValueRef v = LLVMBuildICmp(ctx->ac.builder, pred,
emit_data->args[0], emit_data->args[1],"");
- v = LLVMBuildSExtOrBitCast(builder, v,
- LLVMInt32TypeInContext(context), "");
+ v = LLVMBuildSExtOrBitCast(ctx->ac.builder, v, ctx->i32, "");
emit_data->output[emit_data->chan] = v;
}
struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
- LLVMBuilderRef builder = bld_base->base.gallivm->builder;
-
- LLVMValueRef arg0 = LLVMBuildBitCast(builder, emit_data->args[0],
- bld_base->uint_bld.elem_type, "");
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ LLVMValueRef arg0 = ac_to_integer(&ctx->ac, emit_data->args[0]);
- LLVMValueRef v = LLVMBuildICmp(builder, LLVMIntNE, arg0,
- bld_base->uint_bld.zero, "");
+ LLVMValueRef v = LLVMBuildICmp(ctx->ac.builder, LLVMIntNE, arg0,
+ ctx->i32_0, "");
emit_data->output[emit_data->chan] =
- LLVMBuildSelect(builder, v, emit_data->args[1], emit_data->args[2], "");
+ LLVMBuildSelect(ctx->ac.builder, v, emit_data->args[1], emit_data->args[2], "");
}
static void emit_cmp(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;
+ struct si_shader_context *ctx = si_shader_context(bld_base);
LLVMValueRef cond, *args = emit_data->args;
- cond = LLVMBuildFCmp(builder, LLVMRealOLT, args[0],
- bld_base->base.zero, "");
+ cond = LLVMBuildFCmp(ctx->ac.builder, LLVMRealOLT, args[0],
+ ctx->ac.f32_0, "");
emit_data->output[emit_data->chan] =
- LLVMBuildSelect(builder, cond, args[1], args[2], "");
+ LLVMBuildSelect(ctx->ac.builder, cond, args[1], args[2], "");
}
static void emit_set_cond(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;
+ struct si_shader_context *ctx = si_shader_context(bld_base);
LLVMRealPredicate pred;
LLVMValueRef cond;
default: assert(!"unknown instruction"); pred = 0; break;
}
- cond = LLVMBuildFCmp(builder,
+ cond = LLVMBuildFCmp(ctx->ac.builder,
pred, emit_data->args[0], emit_data->args[1], "");
- emit_data->output[emit_data->chan] = LLVMBuildSelect(builder,
- cond, bld_base->base.one, bld_base->base.zero, "");
+ emit_data->output[emit_data->chan] = LLVMBuildSelect(ctx->ac.builder,
+ cond, ctx->ac.f32_1, ctx->ac.f32_0, "");
}
static void emit_fcmp(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;
- LLVMContextRef context = bld_base->base.gallivm->context;
+ struct si_shader_context *ctx = si_shader_context(bld_base);
LLVMRealPredicate pred;
/* Use ordered for everything but NE (which is usual for
default: assert(!"unknown instruction"); pred = 0; break;
}
- LLVMValueRef v = LLVMBuildFCmp(builder, pred,
+ LLVMValueRef v = LLVMBuildFCmp(ctx->ac.builder, pred,
emit_data->args[0], emit_data->args[1],"");
- v = LLVMBuildSExtOrBitCast(builder, v,
- LLVMInt32TypeInContext(context), "");
+ v = LLVMBuildSExtOrBitCast(ctx->ac.builder, v, ctx->i32, "");
emit_data->output[emit_data->chan] = v;
}
struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
- LLVMBuilderRef builder = bld_base->base.gallivm->builder;
- LLVMContextRef context = bld_base->base.gallivm->context;
+ struct si_shader_context *ctx = si_shader_context(bld_base);
LLVMRealPredicate pred;
/* Use ordered for everything but NE (which is usual for
default: assert(!"unknown instruction"); pred = 0; break;
}
- LLVMValueRef v = LLVMBuildFCmp(builder, pred,
+ LLVMValueRef v = LLVMBuildFCmp(ctx->ac.builder, pred,
emit_data->args[0], emit_data->args[1],"");
- v = LLVMBuildSExtOrBitCast(builder, v,
- LLVMInt32TypeInContext(context), "");
+ v = LLVMBuildSExtOrBitCast(ctx->ac.builder, v, ctx->i32, "");
emit_data->output[emit_data->chan] = v;
}
struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
- LLVMBuilderRef builder = bld_base->base.gallivm->builder;
- LLVMValueRef v = bitcast(bld_base, TGSI_TYPE_UNSIGNED,
- emit_data->args[0]);
- emit_data->output[emit_data->chan] = LLVMBuildNot(builder, v, "");
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ LLVMValueRef v = ac_to_integer(&ctx->ac, emit_data->args[0]);
+ emit_data->output[emit_data->chan] = LLVMBuildNot(ctx->ac.builder, v, "");
}
static void emit_arl(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 floor_index = lp_build_emit_llvm_unary(bld_base, TGSI_OPCODE_FLR, emit_data->args[0]);
- emit_data->output[emit_data->chan] = LLVMBuildFPToSI(builder,
- floor_index, bld_base->base.int_elem_type , "");
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ LLVMValueRef floor_index =
+ ac_build_intrinsic(&ctx->ac, "llvm.floor.f32", ctx->f32,
+ &emit_data->args[0], 1, AC_FUNC_ATTR_READNONE);
+ emit_data->output[emit_data->chan] = LLVMBuildFPToSI(ctx->ac.builder,
+ floor_index, ctx->i32, "");
}
static void emit_and(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;
- emit_data->output[emit_data->chan] = LLVMBuildAnd(builder,
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ emit_data->output[emit_data->chan] = LLVMBuildAnd(ctx->ac.builder,
emit_data->args[0], emit_data->args[1], "");
}
struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
- LLVMBuilderRef builder = bld_base->base.gallivm->builder;
- emit_data->output[emit_data->chan] = LLVMBuildOr(builder,
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ emit_data->output[emit_data->chan] = LLVMBuildOr(ctx->ac.builder,
emit_data->args[0], emit_data->args[1], "");
}
struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
- LLVMBuilderRef builder = bld_base->base.gallivm->builder;
- emit_data->output[emit_data->chan] = LLVMBuildAdd(builder,
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ emit_data->output[emit_data->chan] = LLVMBuildAdd(ctx->ac.builder,
emit_data->args[0], emit_data->args[1], "");
}
struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
- LLVMBuilderRef builder = bld_base->base.gallivm->builder;
- emit_data->output[emit_data->chan] = LLVMBuildUDiv(builder,
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ emit_data->output[emit_data->chan] = LLVMBuildUDiv(ctx->ac.builder,
emit_data->args[0], emit_data->args[1], "");
}
struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
- LLVMBuilderRef builder = bld_base->base.gallivm->builder;
- emit_data->output[emit_data->chan] = LLVMBuildSDiv(builder,
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ emit_data->output[emit_data->chan] = LLVMBuildSDiv(ctx->ac.builder,
emit_data->args[0], emit_data->args[1], "");
}
struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
- LLVMBuilderRef builder = bld_base->base.gallivm->builder;
- emit_data->output[emit_data->chan] = LLVMBuildSRem(builder,
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ emit_data->output[emit_data->chan] = LLVMBuildSRem(ctx->ac.builder,
emit_data->args[0], emit_data->args[1], "");
}
struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
- LLVMBuilderRef builder = bld_base->base.gallivm->builder;
- emit_data->output[emit_data->chan] = LLVMBuildURem(builder,
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ emit_data->output[emit_data->chan] = LLVMBuildURem(ctx->ac.builder,
emit_data->args[0], emit_data->args[1], "");
}
struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
- LLVMBuilderRef builder = bld_base->base.gallivm->builder;
- emit_data->output[emit_data->chan] = LLVMBuildShl(builder,
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ emit_data->output[emit_data->chan] = LLVMBuildShl(ctx->ac.builder,
emit_data->args[0], emit_data->args[1], "");
}
struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
- LLVMBuilderRef builder = bld_base->base.gallivm->builder;
- emit_data->output[emit_data->chan] = LLVMBuildLShr(builder,
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ emit_data->output[emit_data->chan] = LLVMBuildLShr(ctx->ac.builder,
emit_data->args[0], emit_data->args[1], "");
}
static void emit_ishr(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;
- emit_data->output[emit_data->chan] = LLVMBuildAShr(builder,
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ emit_data->output[emit_data->chan] = LLVMBuildAShr(ctx->ac.builder,
emit_data->args[0], emit_data->args[1], "");
}
struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
- LLVMBuilderRef builder = bld_base->base.gallivm->builder;
- emit_data->output[emit_data->chan] = LLVMBuildXor(builder,
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ emit_data->output[emit_data->chan] = LLVMBuildXor(ctx->ac.builder,
emit_data->args[0], emit_data->args[1], "");
}
struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
- LLVMBuilderRef builder = bld_base->base.gallivm->builder;
+ struct si_shader_context *ctx = si_shader_context(bld_base);
- LLVMValueRef cmp, val;
+ LLVMValueRef val;
if (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_I64SSG) {
- cmp = LLVMBuildICmp(builder, LLVMIntSGT, emit_data->args[0], bld_base->int64_bld.zero, "");
- val = LLVMBuildSelect(builder, cmp, bld_base->int64_bld.one, emit_data->args[0], "");
- cmp = LLVMBuildICmp(builder, LLVMIntSGE, val, bld_base->int64_bld.zero, "");
- val = LLVMBuildSelect(builder, cmp, val, LLVMConstInt(bld_base->int64_bld.elem_type, -1, true), "");
+ val = ac_build_isign(&ctx->ac, emit_data->args[0], 64);
} else if (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_ISSG) {
- cmp = LLVMBuildICmp(builder, LLVMIntSGT, emit_data->args[0], bld_base->int_bld.zero, "");
- val = LLVMBuildSelect(builder, cmp, bld_base->int_bld.one, emit_data->args[0], "");
- cmp = LLVMBuildICmp(builder, LLVMIntSGE, val, bld_base->int_bld.zero, "");
- val = LLVMBuildSelect(builder, cmp, val, LLVMConstInt(bld_base->int_bld.elem_type, -1, true), "");
- } else { // float SSG
- cmp = LLVMBuildFCmp(builder, LLVMRealOGT, emit_data->args[0], bld_base->base.zero, "");
- val = LLVMBuildSelect(builder, cmp, bld_base->base.one, emit_data->args[0], "");
- cmp = LLVMBuildFCmp(builder, LLVMRealOGE, val, bld_base->base.zero, "");
- val = LLVMBuildSelect(builder, cmp, val, LLVMConstReal(bld_base->base.elem_type, -1), "");
+ val = ac_build_isign(&ctx->ac, emit_data->args[0], 32);
+ } else if (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_DSSG) {
+ val = ac_build_fsign(&ctx->ac, emit_data->args[0], 64);
+ } else {
+ val = ac_build_fsign(&ctx->ac, emit_data->args[0], 32);
}
emit_data->output[emit_data->chan] = val;
struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
- LLVMBuilderRef builder = bld_base->base.gallivm->builder;
- emit_data->output[emit_data->chan] = LLVMBuildNeg(builder,
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ emit_data->output[emit_data->chan] = LLVMBuildNeg(ctx->ac.builder,
emit_data->args[0], "");
}
struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
- LLVMBuilderRef builder = bld_base->base.gallivm->builder;
- emit_data->output[emit_data->chan] = LLVMBuildFNeg(builder,
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ emit_data->output[emit_data->chan] = LLVMBuildFNeg(ctx->ac.builder,
emit_data->args[0], "");
}
struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
- LLVMBuilderRef builder = bld_base->base.gallivm->builder;
- char *intr;
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ unsigned bitsize;
if (emit_data->info->opcode == TGSI_OPCODE_FRC)
- intr = "llvm.floor.f32";
+ bitsize = 32;
else if (emit_data->info->opcode == TGSI_OPCODE_DFRAC)
- intr = "llvm.floor.f64";
+ bitsize = 64;
else {
assert(0);
return;
}
- LLVMValueRef floor = lp_build_intrinsic(builder, intr, emit_data->dst_type,
- &emit_data->args[0], 1,
- LP_FUNC_ATTR_READNONE);
- emit_data->output[emit_data->chan] = LLVMBuildFSub(builder,
- emit_data->args[0], floor, "");
+ emit_data->output[emit_data->chan] =
+ ac_build_fract(&ctx->ac, emit_data->args[0], bitsize);
}
static void emit_f2i(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;
- emit_data->output[emit_data->chan] = LLVMBuildFPToSI(builder,
- emit_data->args[0], bld_base->int_bld.elem_type, "");
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ emit_data->output[emit_data->chan] = LLVMBuildFPToSI(ctx->ac.builder,
+ emit_data->args[0], ctx->i32, "");
}
static void emit_f2u(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;
- emit_data->output[emit_data->chan] = LLVMBuildFPToUI(builder,
- emit_data->args[0], bld_base->uint_bld.elem_type, "");
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ emit_data->output[emit_data->chan] = LLVMBuildFPToUI(ctx->ac.builder,
+ emit_data->args[0], ctx->i32, "");
}
static void emit_i2f(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;
- emit_data->output[emit_data->chan] = LLVMBuildSIToFP(builder,
- emit_data->args[0], bld_base->base.elem_type, "");
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ emit_data->output[emit_data->chan] = LLVMBuildSIToFP(ctx->ac.builder,
+ emit_data->args[0], ctx->f32, "");
}
static void emit_u2f(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;
- emit_data->output[emit_data->chan] = LLVMBuildUIToFP(builder,
- emit_data->args[0], bld_base->base.elem_type, "");
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ emit_data->output[emit_data->chan] = LLVMBuildUIToFP(ctx->ac.builder,
+ emit_data->args[0], ctx->f32, "");
}
static void
struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
- struct lp_build_context *base = &bld_base->base;
+ struct si_shader_context *ctx = si_shader_context(bld_base);
emit_data->output[emit_data->chan] =
- lp_build_intrinsic(base->gallivm->builder, action->intr_name,
+ ac_build_intrinsic(&ctx->ac, action->intr_name,
emit_data->dst_type, emit_data->args,
- emit_data->arg_count, LP_FUNC_ATTR_READNONE);
+ emit_data->arg_count, AC_FUNC_ATTR_READNONE);
}
static void emit_bfi(const struct lp_build_tgsi_action *action,
struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
- struct gallivm_state *gallivm = bld_base->base.gallivm;
- LLVMBuilderRef builder = gallivm->builder;
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ LLVMBuilderRef builder = ctx->ac.builder;
LLVMValueRef bfi_args[3];
LLVMValueRef bfi_sm5;
LLVMValueRef cond;
bfi_args[0] = LLVMBuildShl(builder,
LLVMBuildSub(builder,
LLVMBuildShl(builder,
- bld_base->int_bld.one,
+ ctx->i32_1,
emit_data->args[3], ""),
- bld_base->int_bld.one, ""),
+ ctx->i32_1, ""),
emit_data->args[2], "");
bfi_args[1] = LLVMBuildShl(builder, emit_data->args[1],
* and disagrees with GLSL semantics when bits (src3) is 32.
*/
cond = LLVMBuildICmp(builder, LLVMIntUGE, emit_data->args[3],
- lp_build_const_int32(gallivm, 32), "");
+ LLVMConstInt(ctx->i32, 32, 0), "");
emit_data->output[emit_data->chan] =
LLVMBuildSelect(builder, cond, emit_data->args[1], bfi_sm5, "");
}
struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
- struct gallivm_state *gallivm = bld_base->base.gallivm;
- LLVMBuilderRef builder = gallivm->builder;
- LLVMValueRef bfe_sm5;
- LLVMValueRef cond;
+ struct si_shader_context *ctx = si_shader_context(bld_base);
- bfe_sm5 = lp_build_intrinsic(builder, action->intr_name,
- emit_data->dst_type, emit_data->args,
- emit_data->arg_count, LP_FUNC_ATTR_READNONE);
+ /* FIXME: LLVM 7 returns incorrect result when count is 0.
+ * https://bugs.freedesktop.org/show_bug.cgi?id=107276
+ */
+ LLVMValueRef zero = ctx->i32_0;
+ LLVMValueRef bfe_sm5 =
+ ac_build_bfe(&ctx->ac, emit_data->args[0],
+ emit_data->args[1], emit_data->args[2],
+ emit_data->info->opcode == TGSI_OPCODE_IBFE);
/* Correct for GLSL semantics. */
- cond = LLVMBuildICmp(builder, LLVMIntUGE, emit_data->args[2],
- lp_build_const_int32(gallivm, 32), "");
+ LLVMValueRef cond = LLVMBuildICmp(ctx->ac.builder, LLVMIntUGE, emit_data->args[2],
+ LLVMConstInt(ctx->i32, 32, 0), "");
+ LLVMValueRef cond2 = LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ, emit_data->args[2],
+ zero, "");
+ bfe_sm5 = LLVMBuildSelect(ctx->ac.builder, cond, emit_data->args[0], bfe_sm5, "");
emit_data->output[emit_data->chan] =
- LLVMBuildSelect(builder, cond, emit_data->args[0], bfe_sm5, "");
+ LLVMBuildSelect(ctx->ac.builder, cond2, zero, bfe_sm5, "");
}
/* this is ffs in C */
struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
- struct gallivm_state *gallivm = bld_base->base.gallivm;
- LLVMBuilderRef builder = gallivm->builder;
- LLVMValueRef args[2] = {
- emit_data->args[0],
-
- /* The value of 1 means that ffs(x=0) = undef, so LLVM won't
- * add special code to check for x=0. The reason is that
- * the LLVM behavior for x=0 is different from what we
- * need here. However, LLVM also assumes that ffs(x) is
- * in [0, 31], but GLSL expects that ffs(0) = -1, so
- * a conditional assignment to handle 0 is still required.
- */
- LLVMConstInt(LLVMInt1TypeInContext(gallivm->context), 1, 0)
- };
-
- LLVMValueRef lsb =
- lp_build_intrinsic(gallivm->builder, "llvm.cttz.i32",
- emit_data->dst_type, args, ARRAY_SIZE(args),
- LP_FUNC_ATTR_READNONE);
+ struct si_shader_context *ctx = si_shader_context(bld_base);
- /* TODO: We need an intrinsic to skip this conditional. */
- /* Check for zero: */
- emit_data->output[emit_data->chan] =
- LLVMBuildSelect(builder,
- LLVMBuildICmp(builder, LLVMIntEQ, args[0],
- bld_base->uint_bld.zero, ""),
- lp_build_const_int32(gallivm, -1), lsb, "");
+ emit_data->output[emit_data->chan] = ac_find_lsb(&ctx->ac, emit_data->dst_type, emit_data->args[0]);
}
/* Find the last bit set. */
struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
- struct gallivm_state *gallivm = bld_base->base.gallivm;
- LLVMBuilderRef builder = gallivm->builder;
- LLVMValueRef args[2] = {
- emit_data->args[0],
- /* Don't generate code for handling zero: */
- LLVMConstInt(LLVMInt1TypeInContext(gallivm->context), 1, 0)
- };
-
- LLVMValueRef msb =
- lp_build_intrinsic(builder, "llvm.ctlz.i32",
- emit_data->dst_type, args, ARRAY_SIZE(args),
- LP_FUNC_ATTR_READNONE);
-
- /* The HW returns the last bit index from MSB, but TGSI wants
- * the index from LSB. Invert it by doing "31 - msb". */
- msb = LLVMBuildSub(builder, lp_build_const_int32(gallivm, 31),
- msb, "");
+ struct si_shader_context *ctx = si_shader_context(bld_base);
- /* Check for zero: */
emit_data->output[emit_data->chan] =
- LLVMBuildSelect(builder,
- LLVMBuildICmp(builder, LLVMIntEQ, args[0],
- bld_base->uint_bld.zero, ""),
- lp_build_const_int32(gallivm, -1), msb, "");
+ ac_build_umsb(&ctx->ac, emit_data->args[0], emit_data->dst_type);
}
/* Find the last bit opposite of the sign bit. */
struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
- struct gallivm_state *gallivm = bld_base->base.gallivm;
- LLVMBuilderRef builder = gallivm->builder;
- LLVMValueRef arg = emit_data->args[0];
-
- LLVMValueRef msb =
- lp_build_intrinsic(builder, "llvm.AMDGPU.flbit.i32",
- emit_data->dst_type, &arg, 1,
- LP_FUNC_ATTR_READNONE);
-
- /* The HW returns the last bit index from MSB, but TGSI wants
- * the index from LSB. Invert it by doing "31 - msb". */
- msb = LLVMBuildSub(builder, lp_build_const_int32(gallivm, 31),
- msb, "");
-
- /* If arg == 0 || arg == -1 (0xffffffff), return -1. */
- LLVMValueRef all_ones = lp_build_const_int32(gallivm, -1);
-
- LLVMValueRef cond =
- LLVMBuildOr(builder,
- LLVMBuildICmp(builder, LLVMIntEQ, arg,
- bld_base->uint_bld.zero, ""),
- LLVMBuildICmp(builder, LLVMIntEQ, arg,
- all_ones, ""), "");
-
+ struct si_shader_context *ctx = si_shader_context(bld_base);
emit_data->output[emit_data->chan] =
- LLVMBuildSelect(builder, cond, all_ones, msb, "");
+ ac_build_imsb(&ctx->ac, emit_data->args[0],
+ emit_data->dst_type);
}
static void emit_iabs(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;
+ struct si_shader_context *ctx = si_shader_context(bld_base);
emit_data->output[emit_data->chan] =
- lp_build_emit_llvm_binary(bld_base, TGSI_OPCODE_IMAX,
- emit_data->args[0],
- LLVMBuildNeg(builder,
- emit_data->args[0], ""));
+ ac_build_imax(&ctx->ac, emit_data->args[0],
+ LLVMBuildNeg(ctx->ac.builder, emit_data->args[0], ""));
}
static void emit_minmax_int(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;
+ struct si_shader_context *ctx = si_shader_context(bld_base);
LLVMIntPredicate op;
switch (emit_data->info->opcode) {
}
emit_data->output[emit_data->chan] =
- LLVMBuildSelect(builder,
- LLVMBuildICmp(builder, op, emit_data->args[0],
+ LLVMBuildSelect(ctx->ac.builder,
+ LLVMBuildICmp(ctx->ac.builder, op, emit_data->args[0],
emit_data->args[1], ""),
emit_data->args[0],
emit_data->args[1], "");
}
-static void pk2h_fetch_args(struct lp_build_tgsi_context *bld_base,
- struct lp_build_emit_data *emit_data)
-{
- emit_data->args[0] = lp_build_emit_fetch(bld_base, emit_data->inst,
- 0, TGSI_CHAN_X);
- emit_data->args[1] = lp_build_emit_fetch(bld_base, emit_data->inst,
- 0, TGSI_CHAN_Y);
-}
-
static void emit_pk2h(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;
- LLVMContextRef context = bld_base->base.gallivm->context;
- struct lp_build_context *uint_bld = &bld_base->uint_bld;
- LLVMTypeRef fp16, i16;
- LLVMValueRef const16, comp[2];
- unsigned i;
-
- fp16 = LLVMHalfTypeInContext(context);
- i16 = LLVMInt16TypeInContext(context);
- const16 = lp_build_const_int32(uint_bld->gallivm, 16);
-
- for (i = 0; i < 2; i++) {
- comp[i] = LLVMBuildFPTrunc(builder, emit_data->args[i], fp16, "");
- comp[i] = LLVMBuildBitCast(builder, comp[i], i16, "");
- comp[i] = LLVMBuildZExt(builder, comp[i], uint_bld->elem_type, "");
- }
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ LLVMValueRef v[] = {
+ lp_build_emit_fetch(bld_base, emit_data->inst, 0, TGSI_CHAN_X),
+ lp_build_emit_fetch(bld_base, emit_data->inst, 0, TGSI_CHAN_Y),
+ };
- comp[1] = LLVMBuildShl(builder, comp[1], const16, "");
- comp[0] = LLVMBuildOr(builder, comp[0], comp[1], "");
- emit_data->output[emit_data->chan] = comp[0];
-}
-
-static void up2h_fetch_args(struct lp_build_tgsi_context *bld_base,
- struct lp_build_emit_data *emit_data)
-{
- emit_data->args[0] = lp_build_emit_fetch(bld_base, emit_data->inst,
- 0, TGSI_CHAN_X);
+ /* From the GLSL 4.50 spec:
+ * "The rounding mode cannot be set and is undefined."
+ *
+ * v_cvt_pkrtz_f16 rounds to zero, but it's fastest.
+ */
+ emit_data->output[emit_data->chan] =
+ LLVMBuildBitCast(ctx->ac.builder, ac_build_cvt_pkrtz_f16(&ctx->ac, v),
+ ctx->i32, "");
}
static void emit_up2h(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;
- LLVMContextRef context = bld_base->base.gallivm->context;
- struct lp_build_context *uint_bld = &bld_base->uint_bld;
- LLVMTypeRef fp16, i16;
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ LLVMTypeRef i16;
LLVMValueRef const16, input, val;
unsigned i;
- fp16 = LLVMHalfTypeInContext(context);
- i16 = LLVMInt16TypeInContext(context);
- const16 = lp_build_const_int32(uint_bld->gallivm, 16);
- input = emit_data->args[0];
+ i16 = LLVMInt16TypeInContext(ctx->ac.context);
+ const16 = LLVMConstInt(ctx->i32, 16, 0);
+ input = lp_build_emit_fetch(bld_base, emit_data->inst, 0, TGSI_CHAN_X);
for (i = 0; i < 2; i++) {
- val = i == 1 ? LLVMBuildLShr(builder, input, const16, "") : input;
- val = LLVMBuildTrunc(builder, val, i16, "");
- val = LLVMBuildBitCast(builder, val, fp16, "");
- emit_data->output[i] =
- LLVMBuildFPExt(builder, val, bld_base->base.elem_type, "");
+ val = i == 1 ? LLVMBuildLShr(ctx->ac.builder, input, const16, "") : input;
+ val = LLVMBuildTrunc(ctx->ac.builder, val, i16, "");
+ val = ac_to_float(&ctx->ac, val);
+ emit_data->output[i] = LLVMBuildFPExt(ctx->ac.builder, val, ctx->f32, "");
}
}
struct si_shader_context *ctx = si_shader_context(bld_base);
emit_data->output[emit_data->chan] =
- LLVMBuildFDiv(bld_base->base.gallivm->builder,
- emit_data->args[0], emit_data->args[1], "");
-
- /* Use v_rcp_f32 instead of precise division. */
- if (HAVE_LLVM >= 0x0309 &&
- !LLVMIsConstant(emit_data->output[emit_data->chan]))
- LLVMSetMetadata(emit_data->output[emit_data->chan],
- ctx->fpmath_md_kind, ctx->fpmath_md_2p5_ulp);
+ ac_build_fdiv(&ctx->ac, emit_data->args[0], emit_data->args[1]);
}
/* 1/sqrt is translated to rsq for f32 if fp32 denormals are not enabled in
struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+
LLVMValueRef sqrt =
- lp_build_emit_llvm_unary(bld_base, TGSI_OPCODE_SQRT,
- emit_data->args[0]);
+ ac_build_intrinsic(&ctx->ac, "llvm.sqrt.f32", ctx->f32,
+ &emit_data->args[0], 1, AC_FUNC_ATTR_READNONE);
+
+ emit_data->output[emit_data->chan] =
+ ac_build_fdiv(&ctx->ac, ctx->ac.f32_1, sqrt);
+}
+
+static void dfracexp_emit(const struct lp_build_tgsi_action *action,
+ struct lp_build_tgsi_context *bld_base,
+ struct lp_build_emit_data *emit_data)
+{
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ LLVMValueRef in = lp_build_emit_fetch(bld_base, emit_data->inst, 0, TGSI_CHAN_X);
emit_data->output[emit_data->chan] =
- lp_build_emit_llvm_binary(bld_base, TGSI_OPCODE_DIV,
- bld_base->base.one, sqrt);
+ ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.frexp.mant.f64",
+ ctx->ac.f64, &in, 1, 0);
+ emit_data->output1[emit_data->chan] =
+ ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.frexp.exp.i32.f64",
+ ctx->ac.i32, &in, 1, 0);
}
-void si_shader_context_init_alu(struct lp_build_tgsi_context *bld_base)
+void si_shader_context_init_alu(struct si_shader_context *ctx)
{
+ struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
+
lp_set_default_actions(bld_base);
bld_base->op_actions[TGSI_OPCODE_AND].emit = emit_and;
bld_base->op_actions[TGSI_OPCODE_ARL].emit = emit_arl;
bld_base->op_actions[TGSI_OPCODE_BFI].emit = emit_bfi;
bld_base->op_actions[TGSI_OPCODE_BREV].emit = build_tgsi_intrinsic_nomem;
- bld_base->op_actions[TGSI_OPCODE_BREV].intr_name =
- HAVE_LLVM >= 0x0308 ? "llvm.bitreverse.i32" : "llvm.AMDGPU.brev";
+ bld_base->op_actions[TGSI_OPCODE_BREV].intr_name = "llvm.bitreverse.i32";
bld_base->op_actions[TGSI_OPCODE_CEIL].emit = build_tgsi_intrinsic_nomem;
bld_base->op_actions[TGSI_OPCODE_CEIL].intr_name = "llvm.ceil.f32";
- bld_base->op_actions[TGSI_OPCODE_CLAMP].emit = build_tgsi_intrinsic_nomem;
- bld_base->op_actions[TGSI_OPCODE_CLAMP].intr_name =
- HAVE_LLVM >= 0x0308 ? "llvm.AMDGPU.clamp." : "llvm.AMDIL.clamp.";
bld_base->op_actions[TGSI_OPCODE_CMP].emit = emit_cmp;
bld_base->op_actions[TGSI_OPCODE_COS].emit = build_tgsi_intrinsic_nomem;
bld_base->op_actions[TGSI_OPCODE_COS].intr_name = "llvm.cos.f32";
bld_base->op_actions[TGSI_OPCODE_DABS].emit = build_tgsi_intrinsic_nomem;
bld_base->op_actions[TGSI_OPCODE_DABS].intr_name = "llvm.fabs.f64";
+ bld_base->op_actions[TGSI_OPCODE_DCEIL].emit = build_tgsi_intrinsic_nomem;
+ bld_base->op_actions[TGSI_OPCODE_DCEIL].intr_name = "llvm.ceil.f64";
+ bld_base->op_actions[TGSI_OPCODE_DFLR].emit = build_tgsi_intrinsic_nomem;
+ bld_base->op_actions[TGSI_OPCODE_DFLR].intr_name = "llvm.floor.f64";
bld_base->op_actions[TGSI_OPCODE_DFMA].emit = build_tgsi_intrinsic_nomem;
bld_base->op_actions[TGSI_OPCODE_DFMA].intr_name = "llvm.fma.f64";
bld_base->op_actions[TGSI_OPCODE_DFRAC].emit = emit_frac;
bld_base->op_actions[TGSI_OPCODE_DIV].emit = emit_fdiv;
bld_base->op_actions[TGSI_OPCODE_DNEG].emit = emit_dneg;
+ bld_base->op_actions[TGSI_OPCODE_DROUND].emit = build_tgsi_intrinsic_nomem;
+ bld_base->op_actions[TGSI_OPCODE_DROUND].intr_name = "llvm.rint.f64";
bld_base->op_actions[TGSI_OPCODE_DSEQ].emit = emit_dcmp;
bld_base->op_actions[TGSI_OPCODE_DSGE].emit = emit_dcmp;
bld_base->op_actions[TGSI_OPCODE_DSLT].emit = emit_dcmp;
bld_base->op_actions[TGSI_OPCODE_DSNE].emit = emit_dcmp;
+ bld_base->op_actions[TGSI_OPCODE_DSSG].emit = emit_ssg;
bld_base->op_actions[TGSI_OPCODE_DRSQ].emit = build_tgsi_intrinsic_nomem;
- bld_base->op_actions[TGSI_OPCODE_DRSQ].intr_name =
- HAVE_LLVM >= 0x0309 ? "llvm.amdgcn.rsq.f64" : "llvm.AMDGPU.rsq.f64";
+ bld_base->op_actions[TGSI_OPCODE_DRSQ].intr_name = "llvm.amdgcn.rsq.f64";
bld_base->op_actions[TGSI_OPCODE_DSQRT].emit = build_tgsi_intrinsic_nomem;
bld_base->op_actions[TGSI_OPCODE_DSQRT].intr_name = "llvm.sqrt.f64";
+ bld_base->op_actions[TGSI_OPCODE_DTRUNC].emit = build_tgsi_intrinsic_nomem;
+ bld_base->op_actions[TGSI_OPCODE_DTRUNC].intr_name = "llvm.trunc.f64";
+ bld_base->op_actions[TGSI_OPCODE_DFRACEXP].emit = dfracexp_emit;
+ bld_base->op_actions[TGSI_OPCODE_DLDEXP].emit = build_tgsi_intrinsic_nomem;
+ bld_base->op_actions[TGSI_OPCODE_DLDEXP].intr_name = "llvm.amdgcn.ldexp.f64";
bld_base->op_actions[TGSI_OPCODE_EX2].emit = build_tgsi_intrinsic_nomem;
- bld_base->op_actions[TGSI_OPCODE_EX2].intr_name =
- HAVE_LLVM >= 0x0308 ? "llvm.exp2.f32" : "llvm.AMDIL.exp.";
+ bld_base->op_actions[TGSI_OPCODE_EX2].intr_name = "llvm.exp2.f32";
bld_base->op_actions[TGSI_OPCODE_FLR].emit = build_tgsi_intrinsic_nomem;
bld_base->op_actions[TGSI_OPCODE_FLR].intr_name = "llvm.floor.f32";
- bld_base->op_actions[TGSI_OPCODE_FMA].emit =
- bld_base->op_actions[TGSI_OPCODE_MAD].emit;
+
+ /* FMA is better on GFX10, because it has FMA units instead of MUL-ADD units. */
+ if (ctx->screen->info.chip_class >= GFX10) {
+ bld_base->op_actions[TGSI_OPCODE_FMA].emit = build_tgsi_intrinsic_nomem;
+ bld_base->op_actions[TGSI_OPCODE_FMA].intr_name = "llvm.fma.f32";
+ } else {
+ bld_base->op_actions[TGSI_OPCODE_FMA].emit =
+ bld_base->op_actions[TGSI_OPCODE_MAD].emit;
+ }
+
bld_base->op_actions[TGSI_OPCODE_FRC].emit = emit_frac;
bld_base->op_actions[TGSI_OPCODE_F2I].emit = emit_f2i;
bld_base->op_actions[TGSI_OPCODE_F2U].emit = emit_f2u;
bld_base->op_actions[TGSI_OPCODE_FSNE].emit = emit_fcmp;
bld_base->op_actions[TGSI_OPCODE_IABS].emit = emit_iabs;
bld_base->op_actions[TGSI_OPCODE_IBFE].emit = emit_bfe;
- bld_base->op_actions[TGSI_OPCODE_IBFE].intr_name = "llvm.AMDGPU.bfe.i32";
bld_base->op_actions[TGSI_OPCODE_IDIV].emit = emit_idiv;
bld_base->op_actions[TGSI_OPCODE_IMAX].emit = emit_minmax_int;
bld_base->op_actions[TGSI_OPCODE_IMIN].emit = emit_minmax_int;
bld_base->op_actions[TGSI_OPCODE_ISLT].emit = emit_icmp;
bld_base->op_actions[TGSI_OPCODE_ISSG].emit = emit_ssg;
bld_base->op_actions[TGSI_OPCODE_I2F].emit = emit_i2f;
- bld_base->op_actions[TGSI_OPCODE_KILL_IF].fetch_args = kill_if_fetch_args;
bld_base->op_actions[TGSI_OPCODE_KILL_IF].emit = kil_emit;
- bld_base->op_actions[TGSI_OPCODE_KILL_IF].intr_name = "llvm.AMDGPU.kill";
- bld_base->op_actions[TGSI_OPCODE_KILL].emit = lp_build_tgsi_intrinsic;
- bld_base->op_actions[TGSI_OPCODE_KILL].intr_name = "llvm.AMDGPU.kilp";
+ bld_base->op_actions[TGSI_OPCODE_KILL].emit = kil_emit;
+ bld_base->op_actions[TGSI_OPCODE_LDEXP].emit = build_tgsi_intrinsic_nomem;
+ bld_base->op_actions[TGSI_OPCODE_LDEXP].intr_name = "llvm.amdgcn.ldexp.f32";
bld_base->op_actions[TGSI_OPCODE_LSB].emit = emit_lsb;
bld_base->op_actions[TGSI_OPCODE_LG2].emit = build_tgsi_intrinsic_nomem;
bld_base->op_actions[TGSI_OPCODE_LG2].intr_name = "llvm.log2.f32";
bld_base->op_actions[TGSI_OPCODE_UMSB].emit = emit_umsb;
bld_base->op_actions[TGSI_OPCODE_NOT].emit = emit_not;
bld_base->op_actions[TGSI_OPCODE_OR].emit = emit_or;
- bld_base->op_actions[TGSI_OPCODE_PK2H].fetch_args = pk2h_fetch_args;
bld_base->op_actions[TGSI_OPCODE_PK2H].emit = emit_pk2h;
bld_base->op_actions[TGSI_OPCODE_POPC].emit = build_tgsi_intrinsic_nomem;
bld_base->op_actions[TGSI_OPCODE_POPC].intr_name = "llvm.ctpop.i32";
bld_base->op_actions[TGSI_OPCODE_TRUNC].intr_name = "llvm.trunc.f32";
bld_base->op_actions[TGSI_OPCODE_UADD].emit = emit_uadd;
bld_base->op_actions[TGSI_OPCODE_UBFE].emit = emit_bfe;
- bld_base->op_actions[TGSI_OPCODE_UBFE].intr_name = "llvm.AMDGPU.bfe.u32";
bld_base->op_actions[TGSI_OPCODE_UDIV].emit = emit_udiv;
bld_base->op_actions[TGSI_OPCODE_UMAX].emit = emit_minmax_int;
bld_base->op_actions[TGSI_OPCODE_UMIN].emit = emit_minmax_int;
bld_base->op_actions[TGSI_OPCODE_U2F].emit = emit_u2f;
bld_base->op_actions[TGSI_OPCODE_XOR].emit = emit_xor;
bld_base->op_actions[TGSI_OPCODE_UCMP].emit = emit_ucmp;
- bld_base->op_actions[TGSI_OPCODE_UP2H].fetch_args = up2h_fetch_args;
bld_base->op_actions[TGSI_OPCODE_UP2H].emit = emit_up2h;
bld_base->op_actions[TGSI_OPCODE_I64MAX].emit = emit_minmax_int;
bld_base->op_actions[TGSI_OPCODE_U64DIV].emit = emit_udiv;
bld_base->op_actions[TGSI_OPCODE_I64DIV].emit = emit_idiv;
}
-
-/* Coordinates for cube map selection. sc, tc, and ma are as in Table 8.27
- * of the OpenGL 4.5 (Compatibility Profile) specification, except ma is
- * already multiplied by two. id is the cube face number.
- */
-struct cube_selection_coords {
- LLVMValueRef stc[2];
- LLVMValueRef ma;
- LLVMValueRef id;
-};
-
-static void build_cube_intrinsic(struct gallivm_state *gallivm,
- LLVMValueRef in[3],
- struct cube_selection_coords *out)
-{
- LLVMBuilderRef builder = gallivm->builder;
-
- if (HAVE_LLVM >= 0x0309) {
- LLVMTypeRef f32 = LLVMTypeOf(in[0]);
-
- out->stc[1] = lp_build_intrinsic(builder, "llvm.amdgcn.cubetc",
- f32, in, 3, LP_FUNC_ATTR_READNONE);
- out->stc[0] = lp_build_intrinsic(builder, "llvm.amdgcn.cubesc",
- f32, in, 3, LP_FUNC_ATTR_READNONE);
- out->ma = lp_build_intrinsic(builder, "llvm.amdgcn.cubema",
- f32, in, 3, LP_FUNC_ATTR_READNONE);
- out->id = lp_build_intrinsic(builder, "llvm.amdgcn.cubeid",
- f32, in, 3, LP_FUNC_ATTR_READNONE);
- } else {
- LLVMValueRef c[4] = {
- in[0],
- in[1],
- in[2],
- LLVMGetUndef(LLVMTypeOf(in[0]))
- };
- LLVMValueRef vec = lp_build_gather_values(gallivm, c, 4);
-
- LLVMValueRef tmp =
- lp_build_intrinsic(builder, "llvm.AMDGPU.cube",
- LLVMTypeOf(vec), &vec, 1,
- LP_FUNC_ATTR_READNONE);
-
- out->stc[1] = LLVMBuildExtractElement(builder, tmp,
- lp_build_const_int32(gallivm, 0), "");
- out->stc[0] = LLVMBuildExtractElement(builder, tmp,
- lp_build_const_int32(gallivm, 1), "");
- out->ma = LLVMBuildExtractElement(builder, tmp,
- lp_build_const_int32(gallivm, 2), "");
- out->id = LLVMBuildExtractElement(builder, tmp,
- lp_build_const_int32(gallivm, 3), "");
- }
-}
-
-/**
- * Build a manual selection sequence for cube face sc/tc coordinates and
- * major axis vector (multiplied by 2 for consistency) for the given
- * vec3 \p coords, for the face implied by \p selcoords.
- *
- * For the major axis, we always adjust the sign to be in the direction of
- * selcoords.ma; i.e., a positive out_ma means that coords is pointed towards
- * the selcoords major axis.
- */
-static void build_cube_select(LLVMBuilderRef builder,
- const struct cube_selection_coords *selcoords,
- const LLVMValueRef *coords,
- LLVMValueRef *out_st,
- LLVMValueRef *out_ma)
-{
- LLVMTypeRef f32 = LLVMTypeOf(coords[0]);
- LLVMValueRef is_ma_positive;
- LLVMValueRef sgn_ma;
- LLVMValueRef is_ma_z, is_not_ma_z;
- LLVMValueRef is_ma_y;
- LLVMValueRef is_ma_x;
- LLVMValueRef sgn;
- LLVMValueRef tmp;
-
- is_ma_positive = LLVMBuildFCmp(builder, LLVMRealUGE,
- selcoords->ma, LLVMConstReal(f32, 0.0), "");
- sgn_ma = LLVMBuildSelect(builder, is_ma_positive,
- LLVMConstReal(f32, 1.0), LLVMConstReal(f32, -1.0), "");
-
- is_ma_z = LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 4.0), "");
- is_not_ma_z = LLVMBuildNot(builder, is_ma_z, "");
- is_ma_y = LLVMBuildAnd(builder, is_not_ma_z,
- LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 2.0), ""), "");
- is_ma_x = LLVMBuildAnd(builder, is_not_ma_z, LLVMBuildNot(builder, is_ma_y, ""), "");
-
- /* Select sc */
- tmp = LLVMBuildSelect(builder, is_ma_z, coords[2], coords[0], "");
- sgn = LLVMBuildSelect(builder, is_ma_y, LLVMConstReal(f32, 1.0),
- LLVMBuildSelect(builder, is_ma_x, sgn_ma,
- LLVMBuildFNeg(builder, sgn_ma, ""), ""), "");
- out_st[0] = LLVMBuildFMul(builder, tmp, sgn, "");
-
- /* Select tc */
- tmp = LLVMBuildSelect(builder, is_ma_y, coords[2], coords[1], "");
- sgn = LLVMBuildSelect(builder, is_ma_y, LLVMBuildFNeg(builder, sgn_ma, ""),
- LLVMConstReal(f32, -1.0), "");
- out_st[1] = LLVMBuildFMul(builder, tmp, sgn, "");
-
- /* Select ma */
- tmp = LLVMBuildSelect(builder, is_ma_z, coords[2],
- LLVMBuildSelect(builder, is_ma_y, coords[1], coords[0], ""), "");
- sgn = LLVMBuildSelect(builder, is_ma_positive,
- LLVMConstReal(f32, 2.0), LLVMConstReal(f32, -2.0), "");
- *out_ma = LLVMBuildFMul(builder, tmp, sgn, "");
-}
-
-void si_prepare_cube_coords(struct lp_build_tgsi_context *bld_base,
- struct lp_build_emit_data *emit_data,
- LLVMValueRef *coords_arg,
- LLVMValueRef *derivs_arg)
-{
-
- unsigned target = emit_data->inst->Texture.Texture;
- unsigned opcode = emit_data->inst->Instruction.Opcode;
- struct gallivm_state *gallivm = bld_base->base.gallivm;
- LLVMBuilderRef builder = gallivm->builder;
- LLVMTypeRef type = bld_base->base.elem_type;
- struct cube_selection_coords selcoords;
- LLVMValueRef coords[4];
- LLVMValueRef invma;
-
- build_cube_intrinsic(gallivm, coords_arg, &selcoords);
-
- invma = lp_build_intrinsic(builder, "llvm.fabs.f32",
- type, &selcoords.ma, 1, LP_FUNC_ATTR_READNONE);
- invma = lp_build_emit_llvm_unary(bld_base, TGSI_OPCODE_RCP, invma);
-
- for (int i = 0; i < 2; ++i)
- coords[i] = LLVMBuildFMul(builder, selcoords.stc[i], invma, "");
-
- coords[2] = selcoords.id;
-
- if (opcode == TGSI_OPCODE_TXD && derivs_arg) {
- LLVMValueRef derivs[4];
- int axis;
-
- /* Convert cube derivatives to 2D derivatives. */
- for (axis = 0; axis < 2; axis++) {
- LLVMValueRef deriv_st[2];
- LLVMValueRef deriv_ma;
-
- /* Transform the derivative alongside the texture
- * coordinate. Mathematically, the correct formula is
- * as follows. Assume we're projecting onto the +Z face
- * and denote by dx/dh the derivative of the (original)
- * X texture coordinate with respect to horizontal
- * window coordinates. The projection onto the +Z face
- * plane is:
- *
- * f(x,z) = x/z
- *
- * Then df/dh = df/dx * dx/dh + df/dz * dz/dh
- * = 1/z * dx/dh - x/z * 1/z * dz/dh.
- *
- * This motivatives the implementation below.
- *
- * Whether this actually gives the expected results for
- * apps that might feed in derivatives obtained via
- * finite differences is anyone's guess. The OpenGL spec
- * seems awfully quiet about how textureGrad for cube
- * maps should be handled.
- */
- build_cube_select(builder, &selcoords, &derivs_arg[axis * 3],
- deriv_st, &deriv_ma);
-
- deriv_ma = LLVMBuildFMul(builder, deriv_ma, invma, "");
-
- for (int i = 0; i < 2; ++i)
- derivs[axis * 2 + i] =
- LLVMBuildFSub(builder,
- LLVMBuildFMul(builder, deriv_st[i], invma, ""),
- LLVMBuildFMul(builder, deriv_ma, coords[i], ""), "");
- }
-
- memcpy(derivs_arg, derivs, sizeof(derivs));
- }
-
- /* Shift the texture coordinate. This must be applied after the
- * derivative calculation.
- */
- for (int i = 0; i < 2; ++i)
- coords[i] = LLVMBuildFAdd(builder, coords[i], LLVMConstReal(type, 1.5), "");
-
- if (target == TGSI_TEXTURE_CUBE_ARRAY ||
- target == TGSI_TEXTURE_SHADOWCUBE_ARRAY) {
- /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
- /* coords_arg.w component - array_index for cube arrays */
- coords[2] = lp_build_emit_llvm_ternary(bld_base, TGSI_OPCODE_MAD,
- coords_arg[3], lp_build_const_float(gallivm, 8.0), coords[2]);
- }
-
- /* Preserve compare/lod/bias. Put it in coords.w. */
- if (opcode == TGSI_OPCODE_TEX2 ||
- opcode == TGSI_OPCODE_TXB2 ||
- opcode == TGSI_OPCODE_TXL2) {
- coords[3] = coords_arg[4];
- } else if (opcode == TGSI_OPCODE_TXB ||
- opcode == TGSI_OPCODE_TXL ||
- target == TGSI_TEXTURE_SHADOWCUBE) {
- coords[3] = coords_arg[3];
- }
-
- memcpy(coords_arg, coords, sizeof(coords));
-}
projects / mesa.git / blobdiff