/*
* Copyright 2017 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"),
static const struct lp_build_tgsi_action tex_action;
-enum desc_type {
- DESC_IMAGE,
- DESC_BUFFER,
- DESC_FMASK,
- DESC_SAMPLER,
-};
-
/**
* Given a v8i32 resource descriptor for a buffer, extract the size of the
* buffer in number of elements and return it as an i32.
LLVMValueRef descriptor)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = &ctx->gallivm;
- LLVMBuilderRef builder = gallivm->builder;
+ LLVMBuilderRef builder = ctx->ac.builder;
LLVMValueRef size =
LLVMBuildExtractElement(builder, descriptor,
LLVMConstInt(ctx->i32, 2, 0), "");
- if (ctx->screen->b.chip_class == VI) {
+ if (ctx->screen->info.chip_class == VI) {
/* On VI, the descriptor contains the size in bytes,
* but TXQ must return the size in elements.
* The stride is always non-zero for resources using TXQ.
static LLVMValueRef
shader_buffer_fetch_rsrc(struct si_shader_context *ctx,
- const struct tgsi_full_src_register *reg)
+ const struct tgsi_full_src_register *reg,
+ bool ubo)
{
LLVMValueRef index;
- LLVMValueRef rsrc_ptr = LLVMGetParam(ctx->main_fn,
- ctx->param_const_and_shader_buffers);
if (!reg->Register.Indirect) {
- index = LLVMConstInt(ctx->i32,
- si_get_shaderbuf_slot(reg->Register.Index), 0);
+ index = LLVMConstInt(ctx->i32, reg->Register.Index, false);
} else {
- index = si_get_bounded_indirect_index(ctx, ®->Indirect,
- reg->Register.Index,
- ctx->num_shader_buffers);
- index = LLVMBuildSub(ctx->gallivm.builder,
- LLVMConstInt(ctx->i32, SI_NUM_SHADER_BUFFERS - 1, 0),
- index, "");
+ index = si_get_indirect_index(ctx, ®->Indirect,
+ 1, reg->Register.Index);
}
- return ac_build_indexed_load_const(&ctx->ac, rsrc_ptr, index);
-}
-
-static bool tgsi_is_array_sampler(unsigned target)
-{
- return target == TGSI_TEXTURE_1D_ARRAY ||
- target == TGSI_TEXTURE_SHADOW1D_ARRAY ||
- target == TGSI_TEXTURE_2D_ARRAY ||
- target == TGSI_TEXTURE_SHADOW2D_ARRAY ||
- target == TGSI_TEXTURE_CUBE_ARRAY ||
- target == TGSI_TEXTURE_SHADOWCUBE_ARRAY ||
- target == TGSI_TEXTURE_2D_ARRAY_MSAA;
+ if (ubo)
+ return ctx->abi.load_ubo(&ctx->abi, index);
+ else
+ return ctx->abi.load_ssbo(&ctx->abi, index, false);
}
static bool tgsi_is_array_image(unsigned target)
static LLVMValueRef force_dcc_off(struct si_shader_context *ctx,
LLVMValueRef rsrc)
{
- if (ctx->screen->b.chip_class <= CIK) {
+ if (ctx->screen->info.chip_class <= CIK) {
return rsrc;
} else {
- LLVMBuilderRef builder = ctx->gallivm.builder;
LLVMValueRef i32_6 = LLVMConstInt(ctx->i32, 6, 0);
LLVMValueRef i32_C = LLVMConstInt(ctx->i32, C_008F28_COMPRESSION_EN, 0);
LLVMValueRef tmp;
- tmp = LLVMBuildExtractElement(builder, rsrc, i32_6, "");
- tmp = LLVMBuildAnd(builder, tmp, i32_C, "");
- return LLVMBuildInsertElement(builder, rsrc, tmp, i32_6, "");
+ tmp = LLVMBuildExtractElement(ctx->ac.builder, rsrc, i32_6, "");
+ tmp = LLVMBuildAnd(ctx->ac.builder, tmp, i32_C, "");
+ return LLVMBuildInsertElement(ctx->ac.builder, rsrc, tmp, i32_6, "");
}
}
-static LLVMValueRef load_image_desc(struct si_shader_context *ctx,
- LLVMValueRef list, LLVMValueRef index,
- unsigned target)
+LLVMValueRef si_load_image_desc(struct si_shader_context *ctx,
+ LLVMValueRef list, LLVMValueRef index,
+ enum ac_descriptor_type desc_type, bool dcc_off)
{
- LLVMBuilderRef builder = ctx->gallivm.builder;
+ LLVMBuilderRef builder = ctx->ac.builder;
+ LLVMValueRef rsrc;
- if (target == TGSI_TEXTURE_BUFFER) {
+ if (desc_type == AC_DESC_BUFFER) {
index = LLVMBuildMul(builder, index,
LLVMConstInt(ctx->i32, 2, 0), "");
index = LLVMBuildAdd(builder, index,
ctx->i32_1, "");
list = LLVMBuildPointerCast(builder, list,
- si_const_array(ctx->v4i32, 0), "");
+ ac_array_in_const32_addr_space(ctx->v4i32), "");
+ } else {
+ assert(desc_type == AC_DESC_IMAGE);
}
- return ac_build_indexed_load_const(&ctx->ac, list, index);
+ rsrc = ac_build_load_to_sgpr(&ctx->ac, list, index);
+ if (desc_type == AC_DESC_IMAGE && dcc_off)
+ rsrc = force_dcc_off(ctx, rsrc);
+ return rsrc;
}
/**
LLVMValueRef index;
bool dcc_off = is_store;
- assert(image->Register.File == TGSI_FILE_IMAGE);
-
if (!image->Register.Indirect) {
const struct tgsi_shader_info *info = bld_base->info;
unsigned images_writemask = info->images_store |
index = si_get_bounded_indirect_index(ctx, &image->Indirect,
image->Register.Index,
ctx->num_images);
- index = LLVMBuildSub(ctx->gallivm.builder,
+ index = LLVMBuildSub(ctx->ac.builder,
LLVMConstInt(ctx->i32, SI_NUM_IMAGES - 1, 0),
index, "");
}
- *rsrc = load_image_desc(ctx, rsrc_ptr, index, target);
- if (dcc_off && target != TGSI_TEXTURE_BUFFER)
- *rsrc = force_dcc_off(ctx, *rsrc);
+ if (image->Register.File != TGSI_FILE_IMAGE) {
+ /* Bindless descriptors are accessible from a different pair of
+ * user SGPR indices.
+ */
+ rsrc_ptr = LLVMGetParam(ctx->main_fn,
+ ctx->param_bindless_samplers_and_images);
+ index = lp_build_emit_fetch_src(bld_base, image,
+ TGSI_TYPE_UNSIGNED, 0);
+
+ /* For simplicity, bindless image descriptors use fixed
+ * 16-dword slots for now.
+ */
+ index = LLVMBuildMul(ctx->ac.builder, index,
+ LLVMConstInt(ctx->i32, 2, 0), "");
+ }
+
+ *rsrc = si_load_image_desc(ctx, rsrc_ptr, index,
+ target == TGSI_TEXTURE_BUFFER ? AC_DESC_BUFFER : AC_DESC_IMAGE,
+ dcc_off);
}
static LLVMValueRef image_fetch_coords(
unsigned src, LLVMValueRef desc)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = &ctx->gallivm;
- LLVMBuilderRef builder = gallivm->builder;
+ LLVMBuilderRef builder = ctx->ac.builder;
unsigned target = inst->Memory.Texture;
unsigned num_coords = tgsi_util_get_texture_coord_dim(target);
LLVMValueRef coords[4];
for (chan = 0; chan < num_coords; ++chan) {
tmp = lp_build_emit_fetch(bld_base, inst, src, chan);
- tmp = LLVMBuildBitCast(builder, tmp, ctx->i32, "");
+ tmp = ac_to_integer(&ctx->ac, tmp);
coords[chan] = tmp;
}
- if (ctx->screen->b.chip_class >= GFX9) {
+ if (ctx->screen->info.chip_class >= GFX9) {
/* 1D textures are allocated and used as 2D on GFX9. */
if (target == TGSI_TEXTURE_1D) {
coords[1] = ctx->i32_0;
num_coords = 4;
}
- return lp_build_gather_values(gallivm, coords, num_coords);
+ return lp_build_gather_values(&ctx->gallivm, coords, num_coords);
}
/**
LLVMValueRef slc = i1false;
LLVMValueRef lwe = i1false;
- if (atomic || (HAVE_LLVM <= 0x0309)) {
+ if (atomic) {
emit_data->args[emit_data->arg_count++] = r128;
emit_data->args[emit_data->arg_count++] = da;
if (!atomic) {
return;
}
- /* HAVE_LLVM >= 0x0400 */
emit_data->args[emit_data->arg_count++] = glc;
emit_data->args[emit_data->arg_count++] = slc;
emit_data->args[emit_data->arg_count++] = lwe;
struct lp_build_emit_data * emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = &ctx->gallivm;
const struct tgsi_full_instruction * inst = emit_data->inst;
unsigned target = inst->Memory.Texture;
LLVMValueRef rsrc;
emit_data->dst_type = ctx->v4f32;
- if (inst->Src[0].Register.File == TGSI_FILE_BUFFER) {
- LLVMBuilderRef builder = gallivm->builder;
+ if (inst->Src[0].Register.File == TGSI_FILE_BUFFER ||
+ inst->Src[0].Register.File == TGSI_FILE_CONSTBUF) {
LLVMValueRef offset;
LLVMValueRef tmp;
- rsrc = shader_buffer_fetch_rsrc(ctx, &inst->Src[0]);
+ bool ubo = inst->Src[0].Register.File == TGSI_FILE_CONSTBUF;
+ rsrc = shader_buffer_fetch_rsrc(ctx, &inst->Src[0], ubo);
tmp = lp_build_emit_fetch(bld_base, inst, 1, 0);
- offset = LLVMBuildBitCast(builder, tmp, ctx->i32, "");
+ offset = ac_to_integer(&ctx->ac, tmp);
buffer_append_args(ctx, emit_data, rsrc, ctx->i32_0,
offset, false, false);
- } else if (inst->Src[0].Register.File == TGSI_FILE_IMAGE) {
+ } else if (inst->Src[0].Register.File == TGSI_FILE_IMAGE ||
+ tgsi_is_bindless_image_file(inst->Src[0].Register.File)) {
LLVMValueRef coords;
image_fetch_rsrc(bld_base, &inst->Src[0], false, target, &rsrc);
}
}
-static unsigned get_load_intr_attribs(bool readonly_memory)
-{
- /* READNONE means writes can't affect it, while READONLY means that
- * writes can affect it. */
- return readonly_memory && HAVE_LLVM >= 0x0400 ?
- LP_FUNC_ATTR_READNONE :
- LP_FUNC_ATTR_READONLY;
-}
-
-static unsigned get_store_intr_attribs(bool writeonly_memory)
-{
- return writeonly_memory && HAVE_LLVM >= 0x0400 ?
- LP_FUNC_ATTR_INACCESSIBLE_MEM_ONLY :
- LP_FUNC_ATTR_WRITEONLY;
-}
-
static void load_emit_buffer(struct si_shader_context *ctx,
struct lp_build_emit_data *emit_data,
- bool readonly_memory)
+ bool can_speculate, bool allow_smem)
{
const struct tgsi_full_instruction *inst = emit_data->inst;
- struct gallivm_state *gallivm = &ctx->gallivm;
- LLVMBuilderRef builder = gallivm->builder;
uint writemask = inst->Dst[0].Register.WriteMask;
uint count = util_last_bit(writemask);
- const char *intrinsic_name;
- LLVMTypeRef dst_type;
+ LLVMValueRef *args = emit_data->args;
- switch (count) {
- case 1:
- intrinsic_name = "llvm.amdgcn.buffer.load.f32";
- dst_type = ctx->f32;
- break;
- case 2:
- intrinsic_name = "llvm.amdgcn.buffer.load.v2f32";
- dst_type = LLVMVectorType(ctx->f32, 2);
- break;
- default: // 3 & 4
- intrinsic_name = "llvm.amdgcn.buffer.load.v4f32";
- dst_type = ctx->v4f32;
- count = 4;
- }
+ /* Don't use SMEM for shader buffer loads, because LLVM doesn't
+ * select SMEM for SI.load.const with a non-constant offset, and
+ * constant offsets practically don't exist with shader buffers.
+ *
+ * Also, SI.load.const doesn't use inst_offset when it's lowered
+ * to VMEM, so we just end up with more VALU instructions in the end
+ * and no benefit.
+ *
+ * TODO: Remove this line once LLVM can select SMEM with a non-constant
+ * offset, and can derive inst_offset when VMEM is selected.
+ * After that, si_memory_barrier should invalidate sL1 for shader
+ * buffers.
+ */
- emit_data->output[emit_data->chan] = lp_build_intrinsic(
- builder, intrinsic_name, dst_type,
- emit_data->args, emit_data->arg_count,
- get_load_intr_attribs(readonly_memory));
+ assert(LLVMConstIntGetZExtValue(args[1]) == 0); /* vindex */
+ emit_data->output[emit_data->chan] =
+ ac_build_buffer_load(&ctx->ac, args[0], count, NULL,
+ args[2], NULL, 0,
+ LLVMConstIntGetZExtValue(args[3]),
+ LLVMConstIntGetZExtValue(args[4]),
+ can_speculate, allow_smem);
}
static LLVMValueRef get_memory_ptr(struct si_shader_context *ctx,
const struct tgsi_full_instruction *inst,
LLVMTypeRef type, int arg)
{
- struct gallivm_state *gallivm = &ctx->gallivm;
- LLVMBuilderRef builder = gallivm->builder;
+ LLVMBuilderRef builder = ctx->ac.builder;
LLVMValueRef offset, ptr;
int addr_space;
offset = lp_build_emit_fetch(&ctx->bld_base, inst, arg, 0);
- offset = LLVMBuildBitCast(builder, offset, ctx->i32, "");
+ offset = ac_to_integer(&ctx->ac, offset);
- ptr = ctx->shared_memory;
+ ptr = ctx->ac.lds;
ptr = LLVMBuildGEP(builder, ptr, &offset, 1, "");
addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
ptr = LLVMBuildBitCast(builder, ptr, LLVMPointerType(type, addr_space), "");
struct lp_build_emit_data *emit_data)
{
const struct tgsi_full_instruction *inst = emit_data->inst;
- struct gallivm_state *gallivm = &ctx->gallivm;
- LLVMBuilderRef builder = gallivm->builder;
unsigned writemask = inst->Dst[0].Register.WriteMask;
LLVMValueRef channels[4], ptr, derived_ptr, index;
int chan;
}
index = LLVMConstInt(ctx->i32, chan, 0);
- derived_ptr = LLVMBuildGEP(builder, ptr, &index, 1, "");
- channels[chan] = LLVMBuildLoad(builder, derived_ptr, "");
+ derived_ptr = LLVMBuildGEP(ctx->ac.builder, ptr, &index, 1, "");
+ channels[chan] = LLVMBuildLoad(ctx->ac.builder, derived_ptr, "");
}
- emit_data->output[emit_data->chan] = lp_build_gather_values(gallivm, channels, 4);
+ emit_data->output[emit_data->chan] = lp_build_gather_values(&ctx->gallivm, channels, 4);
}
/**
* images.
*/
if (inst->Src[0].Register.File == TGSI_FILE_BUFFER ||
- (inst->Src[0].Register.File == TGSI_FILE_IMAGE &&
- inst->Memory.Texture == TGSI_TEXTURE_BUFFER)) {
+ (inst->Memory.Texture == TGSI_TEXTURE_BUFFER &&
+ (inst->Src[0].Register.File == TGSI_FILE_IMAGE ||
+ tgsi_is_bindless_image_file(inst->Src[0].Register.File)))) {
if (!shader_buffers_reverse_access_mask &&
!(info->images_buffers & images_reverse_access_mask))
return true;
struct lp_build_emit_data *emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = &ctx->gallivm;
- LLVMBuilderRef builder = gallivm->builder;
+ LLVMBuilderRef builder = ctx->ac.builder;
const struct tgsi_full_instruction * inst = emit_data->inst;
const struct tgsi_shader_info *info = &ctx->shader->selector->info;
char intrinsic_name[64];
- bool readonly_memory = false;
+ bool can_speculate = false;
if (inst->Src[0].Register.File == TGSI_FILE_MEMORY) {
load_emit_memory(ctx, emit_data);
return;
}
+ if (inst->Src[0].Register.File == TGSI_FILE_CONSTBUF) {
+ load_emit_buffer(ctx, emit_data, true, true);
+ return;
+ }
+
if (inst->Memory.Qualifier & TGSI_MEMORY_VOLATILE)
- si_emit_waitcnt(ctx, VM_CNT);
+ ac_build_waitcnt(&ctx->ac, VM_CNT);
- readonly_memory = !(inst->Memory.Qualifier & TGSI_MEMORY_VOLATILE) &&
+ can_speculate = !(inst->Memory.Qualifier & TGSI_MEMORY_VOLATILE) &&
is_oneway_access_only(inst, info,
info->shader_buffers_store |
info->shader_buffers_atomic,
info->images_atomic);
if (inst->Src[0].Register.File == TGSI_FILE_BUFFER) {
- load_emit_buffer(ctx, emit_data, readonly_memory);
+ load_emit_buffer(ctx, emit_data, can_speculate, false);
return;
}
if (inst->Memory.Texture == TGSI_TEXTURE_BUFFER) {
+ unsigned num_channels = util_last_bit(inst->Dst[0].Register.WriteMask);
+ LLVMValueRef result =
+ ac_build_buffer_load_format(&ctx->ac,
+ emit_data->args[0],
+ emit_data->args[1],
+ emit_data->args[2],
+ num_channels,
+ LLVMConstIntGetZExtValue(emit_data->args[3]),
+ can_speculate);
emit_data->output[emit_data->chan] =
- lp_build_intrinsic(
- builder, "llvm.amdgcn.buffer.load.format.v4f32", emit_data->dst_type,
- emit_data->args, emit_data->arg_count,
- get_load_intr_attribs(readonly_memory));
+ ac_build_expand_to_vec4(&ctx->ac, result, num_channels);
} else {
ac_get_image_intr_name("llvm.amdgcn.image.load",
emit_data->dst_type, /* vdata */
lp_build_intrinsic(
builder, intrinsic_name, emit_data->dst_type,
emit_data->args, emit_data->arg_count,
- get_load_intr_attribs(readonly_memory));
+ ac_get_load_intr_attribs(can_speculate));
}
}
struct lp_build_emit_data * emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = &ctx->gallivm;
- LLVMBuilderRef builder = gallivm->builder;
const struct tgsi_full_instruction * inst = emit_data->inst;
struct tgsi_full_src_register memory;
LLVMValueRef chans[4];
LLVMValueRef rsrc;
unsigned chan;
- emit_data->dst_type = LLVMVoidTypeInContext(gallivm->context);
+ emit_data->dst_type = ctx->voidt;
for (chan = 0; chan < 4; ++chan) {
chans[chan] = lp_build_emit_fetch(bld_base, inst, 1, chan);
}
- data = lp_build_gather_values(gallivm, chans, 4);
+ data = lp_build_gather_values(&ctx->gallivm, chans, 4);
emit_data->args[emit_data->arg_count++] = data;
LLVMValueRef offset;
LLVMValueRef tmp;
- rsrc = shader_buffer_fetch_rsrc(ctx, &memory);
+ rsrc = shader_buffer_fetch_rsrc(ctx, &memory, false);
tmp = lp_build_emit_fetch(bld_base, inst, 0, 0);
- offset = LLVMBuildBitCast(builder, tmp, ctx->i32, "");
+ offset = ac_to_integer(&ctx->ac, tmp);
buffer_append_args(ctx, emit_data, rsrc, ctx->i32_0,
offset, false, false);
- } else if (inst->Dst[0].Register.File == TGSI_FILE_IMAGE) {
+ } else if (inst->Dst[0].Register.File == TGSI_FILE_IMAGE ||
+ tgsi_is_bindless_image_file(inst->Dst[0].Register.File)) {
unsigned target = inst->Memory.Texture;
LLVMValueRef coords;
* The only way to get unaligned stores in radeonsi is through
* shader images.
*/
- bool force_glc = ctx->screen->b.chip_class == SI;
+ bool force_glc = ctx->screen->info.chip_class == SI;
image_fetch_rsrc(bld_base, &memory, true, target, &rsrc);
coords = image_fetch_coords(bld_base, inst, 0, rsrc);
bool writeonly_memory)
{
const struct tgsi_full_instruction *inst = emit_data->inst;
- struct gallivm_state *gallivm = &ctx->gallivm;
- LLVMBuilderRef builder = gallivm->builder;
+ LLVMBuilderRef builder = ctx->ac.builder;
LLVMValueRef base_data = emit_data->args[0];
LLVMValueRef base_offset = emit_data->args[3];
unsigned writemask = inst->Dst[0].Register.WriteMask;
lp_build_intrinsic(
builder, intrinsic_name, emit_data->dst_type,
emit_data->args, emit_data->arg_count,
- get_store_intr_attribs(writeonly_memory));
+ ac_get_store_intr_attribs(writeonly_memory));
}
}
struct lp_build_emit_data *emit_data)
{
const struct tgsi_full_instruction *inst = emit_data->inst;
- struct gallivm_state *gallivm = &ctx->gallivm;
- LLVMBuilderRef builder = gallivm->builder;
+ LLVMBuilderRef builder = ctx->ac.builder;
unsigned writemask = inst->Dst[0].Register.WriteMask;
LLVMValueRef ptr, derived_ptr, data, index;
int chan;
struct lp_build_emit_data *emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = &ctx->gallivm;
- LLVMBuilderRef builder = gallivm->builder;
+ LLVMBuilderRef builder = ctx->ac.builder;
const struct tgsi_full_instruction * inst = emit_data->inst;
const struct tgsi_shader_info *info = &ctx->shader->selector->info;
unsigned target = inst->Memory.Texture;
}
if (inst->Memory.Qualifier & TGSI_MEMORY_VOLATILE)
- si_emit_waitcnt(ctx, VM_CNT);
+ ac_build_waitcnt(&ctx->ac, VM_CNT);
writeonly_memory = is_oneway_access_only(inst, info,
info->shader_buffers_load |
builder, "llvm.amdgcn.buffer.store.format.v4f32",
emit_data->dst_type, emit_data->args,
emit_data->arg_count,
- get_store_intr_attribs(writeonly_memory));
+ ac_get_store_intr_attribs(writeonly_memory));
} else {
ac_get_image_intr_name("llvm.amdgcn.image.store",
LLVMTypeOf(emit_data->args[0]), /* vdata */
lp_build_intrinsic(
builder, intrinsic_name, emit_data->dst_type,
emit_data->args, emit_data->arg_count,
- get_store_intr_attribs(writeonly_memory));
+ ac_get_store_intr_attribs(writeonly_memory));
}
}
struct lp_build_emit_data * emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = &ctx->gallivm;
- LLVMBuilderRef builder = gallivm->builder;
const struct tgsi_full_instruction * inst = emit_data->inst;
LLVMValueRef data1, data2;
LLVMValueRef rsrc;
emit_data->dst_type = ctx->f32;
tmp = lp_build_emit_fetch(bld_base, inst, 2, 0);
- data1 = LLVMBuildBitCast(builder, tmp, ctx->i32, "");
+ data1 = ac_to_integer(&ctx->ac, tmp);
if (inst->Instruction.Opcode == TGSI_OPCODE_ATOMCAS) {
tmp = lp_build_emit_fetch(bld_base, inst, 3, 0);
- data2 = LLVMBuildBitCast(builder, tmp, ctx->i32, "");
+ data2 = ac_to_integer(&ctx->ac, tmp);
}
/* llvm.amdgcn.image/buffer.atomic.cmpswap reflect the hardware order
if (inst->Src[0].Register.File == TGSI_FILE_BUFFER) {
LLVMValueRef offset;
- rsrc = shader_buffer_fetch_rsrc(ctx, &inst->Src[0]);
+ rsrc = shader_buffer_fetch_rsrc(ctx, &inst->Src[0], false);
tmp = lp_build_emit_fetch(bld_base, inst, 1, 0);
- offset = LLVMBuildBitCast(builder, tmp, ctx->i32, "");
+ offset = ac_to_integer(&ctx->ac, tmp);
buffer_append_args(ctx, emit_data, rsrc, ctx->i32_0,
offset, true, false);
- } else if (inst->Src[0].Register.File == TGSI_FILE_IMAGE) {
+ } else if (inst->Src[0].Register.File == TGSI_FILE_IMAGE ||
+ tgsi_is_bindless_image_file(inst->Src[0].Register.File)) {
unsigned target = inst->Memory.Texture;
LLVMValueRef coords;
static void atomic_emit_memory(struct si_shader_context *ctx,
struct lp_build_emit_data *emit_data) {
- struct gallivm_state *gallivm = &ctx->gallivm;
- LLVMBuilderRef builder = gallivm->builder;
+ LLVMBuilderRef builder = ctx->ac.builder;
const struct tgsi_full_instruction * inst = emit_data->inst;
LLVMValueRef ptr, result, arg;
ptr = get_memory_ptr(ctx, inst, ctx->i32, 1);
arg = lp_build_emit_fetch(&ctx->bld_base, inst, 2, 0);
- arg = LLVMBuildBitCast(builder, arg, ctx->i32, "");
+ arg = ac_to_integer(&ctx->ac, arg);
if (inst->Instruction.Opcode == TGSI_OPCODE_ATOMCAS) {
LLVMValueRef new_data;
new_data = lp_build_emit_fetch(&ctx->bld_base,
inst, 3, 0);
- new_data = LLVMBuildBitCast(builder, new_data, ctx->i32, "");
+ new_data = ac_to_integer(&ctx->ac, new_data);
result = LLVMBuildAtomicCmpXchg(builder, ptr, arg, new_data,
LLVMAtomicOrderingSequentiallyConsistent,
struct lp_build_emit_data *emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = &ctx->gallivm;
- LLVMBuilderRef builder = gallivm->builder;
+ LLVMBuilderRef builder = ctx->ac.builder;
const struct tgsi_full_instruction * inst = emit_data->inst;
char intrinsic_name[40];
LLVMValueRef tmp;
tmp = lp_build_intrinsic(
builder, intrinsic_name, ctx->i32,
emit_data->args, emit_data->arg_count, 0);
- emit_data->output[emit_data->chan] =
- LLVMBuildBitCast(builder, tmp, ctx->f32, "");
+ emit_data->output[emit_data->chan] = ac_to_float(&ctx->ac, tmp);
}
static void set_tex_fetch_args(struct si_shader_context *ctx,
LLVMValueRef *param, unsigned count,
unsigned dmask)
{
- struct gallivm_state *gallivm = &ctx->gallivm;
struct ac_image_args args = {};
/* Pad to power of two vector */
param[count++] = LLVMGetUndef(ctx->i32);
if (count > 1)
- args.addr = lp_build_gather_values(gallivm, param, count);
+ args.addr = lp_build_gather_values(&ctx->gallivm, param, count);
else
args.addr = param[0];
static LLVMValueRef fix_resinfo(struct si_shader_context *ctx,
unsigned target, LLVMValueRef out)
{
- LLVMBuilderRef builder = ctx->gallivm.builder;
+ LLVMBuilderRef builder = ctx->ac.builder;
/* 1D textures are allocated and used as 2D on GFX9. */
- if (ctx->screen->b.chip_class >= GFX9 &&
+ if (ctx->screen->info.chip_class >= GFX9 &&
(target == TGSI_TEXTURE_1D_ARRAY ||
target == TGSI_TEXTURE_SHADOW1D_ARRAY)) {
LLVMValueRef layers =
emit_data->dst_type = ctx->v4i32;
if (reg->Register.File == TGSI_FILE_BUFFER) {
- emit_data->args[0] = shader_buffer_fetch_rsrc(ctx, reg);
+ emit_data->args[0] = shader_buffer_fetch_rsrc(ctx, reg, false);
emit_data->arg_count = 1;
} else if (inst->Memory.Texture == TGSI_TEXTURE_BUFFER) {
image_fetch_rsrc(bld_base, reg, false, inst->Memory.Texture,
struct lp_build_emit_data *emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = &ctx->gallivm;
- LLVMBuilderRef builder = gallivm->builder;
+ LLVMBuilderRef builder = ctx->ac.builder;
const struct tgsi_full_instruction *inst = emit_data->inst;
LLVMValueRef out;
/**
* Load an image view, fmask view. or sampler state descriptor.
*/
-static LLVMValueRef load_sampler_desc(struct si_shader_context *ctx,
- LLVMValueRef list, LLVMValueRef index,
- enum desc_type type)
+LLVMValueRef si_load_sampler_desc(struct si_shader_context *ctx,
+ LLVMValueRef list, LLVMValueRef index,
+ enum ac_descriptor_type type)
{
- struct gallivm_state *gallivm = &ctx->gallivm;
- LLVMBuilderRef builder = gallivm->builder;
+ LLVMBuilderRef builder = ctx->ac.builder;
switch (type) {
- case DESC_IMAGE:
+ case AC_DESC_IMAGE:
/* The image is at [0:7]. */
index = LLVMBuildMul(builder, index, LLVMConstInt(ctx->i32, 2, 0), "");
break;
- case DESC_BUFFER:
+ case AC_DESC_BUFFER:
/* The buffer is in [4:7]. */
index = LLVMBuildMul(builder, index, LLVMConstInt(ctx->i32, 4, 0), "");
index = LLVMBuildAdd(builder, index, ctx->i32_1, "");
list = LLVMBuildPointerCast(builder, list,
- si_const_array(ctx->v4i32, 0), "");
+ ac_array_in_const32_addr_space(ctx->v4i32), "");
break;
- case DESC_FMASK:
+ case AC_DESC_FMASK:
/* The FMASK is at [8:15]. */
index = LLVMBuildMul(builder, index, LLVMConstInt(ctx->i32, 2, 0), "");
index = LLVMBuildAdd(builder, index, ctx->i32_1, "");
break;
- case DESC_SAMPLER:
+ case AC_DESC_SAMPLER:
/* The sampler state is at [12:15]. */
index = LLVMBuildMul(builder, index, LLVMConstInt(ctx->i32, 4, 0), "");
index = LLVMBuildAdd(builder, index, LLVMConstInt(ctx->i32, 3, 0), "");
list = LLVMBuildPointerCast(builder, list,
- si_const_array(ctx->v4i32, 0), "");
+ ac_array_in_const32_addr_space(ctx->v4i32), "");
break;
}
- return ac_build_indexed_load_const(&ctx->ac, list, index);
+ return ac_build_load_to_sgpr(&ctx->ac, list, index);
}
/* Disable anisotropic filtering if BASE_LEVEL == LAST_LEVEL.
static LLVMValueRef sici_fix_sampler_aniso(struct si_shader_context *ctx,
LLVMValueRef res, LLVMValueRef samp)
{
- LLVMBuilderRef builder = ctx->gallivm.builder;
LLVMValueRef img7, samp0;
- if (ctx->screen->b.chip_class >= VI)
+ if (ctx->screen->info.chip_class >= VI)
return samp;
- img7 = LLVMBuildExtractElement(builder, res,
+ img7 = LLVMBuildExtractElement(ctx->ac.builder, res,
LLVMConstInt(ctx->i32, 7, 0), "");
- samp0 = LLVMBuildExtractElement(builder, samp,
+ samp0 = LLVMBuildExtractElement(ctx->ac.builder, samp,
ctx->i32_0, "");
- samp0 = LLVMBuildAnd(builder, samp0, img7, "");
- return LLVMBuildInsertElement(builder, samp, samp0,
+ samp0 = LLVMBuildAnd(ctx->ac.builder, samp0, img7, "");
+ return LLVMBuildInsertElement(ctx->ac.builder, samp, samp0,
ctx->i32_0, "");
}
®->Indirect,
reg->Register.Index,
ctx->num_samplers);
- index = LLVMBuildAdd(ctx->gallivm.builder, index,
+ index = LLVMBuildAdd(ctx->ac.builder, index,
LLVMConstInt(ctx->i32, SI_NUM_IMAGES / 2, 0), "");
} else {
index = LLVMConstInt(ctx->i32,
si_get_sampler_slot(reg->Register.Index), 0);
}
+ if (reg->Register.File != TGSI_FILE_SAMPLER) {
+ /* Bindless descriptors are accessible from a different pair of
+ * user SGPR indices.
+ */
+ list = LLVMGetParam(ctx->main_fn,
+ ctx->param_bindless_samplers_and_images);
+ index = lp_build_emit_fetch_src(bld_base, reg,
+ TGSI_TYPE_UNSIGNED, 0);
+ }
+
if (target == TGSI_TEXTURE_BUFFER)
- *res_ptr = load_sampler_desc(ctx, list, index, DESC_BUFFER);
+ *res_ptr = si_load_sampler_desc(ctx, list, index, AC_DESC_BUFFER);
else
- *res_ptr = load_sampler_desc(ctx, list, index, DESC_IMAGE);
+ *res_ptr = si_load_sampler_desc(ctx, list, index, AC_DESC_IMAGE);
if (samp_ptr)
*samp_ptr = NULL;
if (target == TGSI_TEXTURE_2D_MSAA ||
target == TGSI_TEXTURE_2D_ARRAY_MSAA) {
if (fmask_ptr)
- *fmask_ptr = load_sampler_desc(ctx, list, index,
- DESC_FMASK);
+ *fmask_ptr = si_load_sampler_desc(ctx, list, index,
+ AC_DESC_FMASK);
} else if (target != TGSI_TEXTURE_BUFFER) {
if (samp_ptr) {
- *samp_ptr = load_sampler_desc(ctx, list, index,
- DESC_SAMPLER);
+ *samp_ptr = si_load_sampler_desc(ctx, list, index,
+ AC_DESC_SAMPLER);
*samp_ptr = sici_fix_sampler_aniso(ctx, *res_ptr, *samp_ptr);
}
}
struct lp_build_emit_data *emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = &ctx->gallivm;
const struct tgsi_full_instruction *inst = emit_data->inst;
unsigned opcode = inst->Instruction.Opcode;
unsigned target = inst->Texture.Texture;
/* Fetch and project texture coordinates */
coords[3] = lp_build_emit_fetch(bld_base, emit_data->inst, 0, TGSI_CHAN_W);
- for (chan = 0; chan < 3; chan++ ) {
+ for (chan = 0; chan < 3; chan++) {
coords[chan] = lp_build_emit_fetch(bld_base,
emit_data->inst, 0,
chan);
}
if (opcode == TGSI_OPCODE_TXP)
- coords[3] = bld_base->base.one;
+ coords[3] = ctx->ac.f32_1;
/* Pack offsets. */
if (has_offset &&
for (chan = 0; chan < 3; chan++) {
offset[chan] = lp_build_emit_fetch_texoffset(bld_base,
emit_data->inst, 0, chan);
- offset[chan] = LLVMBuildAnd(gallivm->builder, offset[chan],
+ offset[chan] = LLVMBuildAnd(ctx->ac.builder, offset[chan],
LLVMConstInt(ctx->i32, 0x3f, 0), "");
if (chan)
- offset[chan] = LLVMBuildShl(gallivm->builder, offset[chan],
+ offset[chan] = LLVMBuildShl(ctx->ac.builder, offset[chan],
LLVMConstInt(ctx->i32, chan*8, 0), "");
}
- pack = LLVMBuildOr(gallivm->builder, offset[0], offset[1], "");
- pack = LLVMBuildOr(gallivm->builder, pack, offset[2], "");
+ pack = LLVMBuildOr(ctx->ac.builder, offset[0], offset[1], "");
+ pack = LLVMBuildOr(ctx->ac.builder, pack, offset[2], "");
address[count++] = pack;
}
z = coords[ref_pos];
}
- /* TC-compatible HTILE promotes Z16 and Z24 to Z32_FLOAT,
+ /* Section 8.23.1 (Depth Texture Comparison Mode) of the
+ * OpenGL 4.5 spec says:
+ *
+ * "If the texture’s internal format indicates a fixed-point
+ * depth texture, then D_t and D_ref are clamped to the
+ * range [0, 1]; otherwise no clamping is performed."
+ *
+ * TC-compatible HTILE promotes Z16 and Z24 to Z32_FLOAT,
* so the depth comparison value isn't clamped for Z16 and
* Z24 anymore. Do it manually here.
- *
- * It's unnecessary if the original texture format was
- * Z32_FLOAT, but we don't know that here.
*/
- if (ctx->screen->b.chip_class == VI)
- z = ac_build_clamp(&ctx->ac, z);
+ if (ctx->screen->info.chip_class >= VI) {
+ LLVMValueRef upgraded;
+ LLVMValueRef clamped;
+ upgraded = LLVMBuildExtractElement(ctx->ac.builder, samp_ptr,
+ LLVMConstInt(ctx->i32, 3, false), "");
+ upgraded = LLVMBuildLShr(ctx->ac.builder, upgraded,
+ LLVMConstInt(ctx->i32, 29, false), "");
+ upgraded = LLVMBuildTrunc(ctx->ac.builder, upgraded, ctx->i1, "");
+ clamped = ac_build_clamp(&ctx->ac, z);
+ z = LLVMBuildSelect(ctx->ac.builder, upgraded, clamped, z, "");
+ }
address[count++] = z;
}
num_src_deriv_channels = 1;
/* 1D textures are allocated and used as 2D on GFX9. */
- if (ctx->screen->b.chip_class >= GFX9) {
+ if (ctx->screen->info.chip_class >= GFX9) {
num_dst_deriv_channels = 2;
num_deriv_channels = 2;
} else {
for (chan = num_src_deriv_channels;
chan < num_dst_deriv_channels; chan++)
derivs[param * num_dst_deriv_channels + chan] =
- bld_base->base.zero;
+ ctx->ac.f32_0;
}
}
if (target == TGSI_TEXTURE_CUBE ||
target == TGSI_TEXTURE_CUBE_ARRAY ||
target == TGSI_TEXTURE_SHADOWCUBE ||
- target == TGSI_TEXTURE_SHADOWCUBE_ARRAY)
+ target == TGSI_TEXTURE_SHADOWCUBE_ARRAY) {
ac_prepare_cube_coords(&ctx->ac,
opcode == TGSI_OPCODE_TXD,
target == TGSI_TEXTURE_CUBE_ARRAY ||
target == TGSI_TEXTURE_SHADOWCUBE_ARRAY,
+ opcode == TGSI_OPCODE_LODQ,
coords, derivs);
+ } else if (tgsi_is_array_sampler(target) &&
+ opcode != TGSI_OPCODE_TXF &&
+ opcode != TGSI_OPCODE_TXF_LZ &&
+ ctx->screen->info.chip_class <= VI) {
+ unsigned array_coord = target == TGSI_TEXTURE_1D_ARRAY ? 1 : 2;
+ coords[array_coord] =
+ ac_build_intrinsic(&ctx->ac, "llvm.rint.f32", ctx->f32,
+ &coords[array_coord], 1, 0);
+ }
if (opcode == TGSI_OPCODE_TXD)
for (int i = 0; i < num_deriv_channels * 2; i++)
address[count++] = coords[2];
/* 1D textures are allocated and used as 2D on GFX9. */
- if (ctx->screen->b.chip_class >= GFX9) {
+ if (ctx->screen->info.chip_class >= GFX9) {
LLVMValueRef filler;
/* Use 0.5, so that we don't sample the border color. */
- if (opcode == TGSI_OPCODE_TXF)
+ if (opcode == TGSI_OPCODE_TXF ||
+ opcode == TGSI_OPCODE_TXF_LZ)
filler = ctx->i32_0;
else
filler = LLVMConstReal(ctx->f32, 0.5);
count = 16;
}
- for (chan = 0; chan < count; chan++ ) {
- address[chan] = LLVMBuildBitCast(gallivm->builder,
- address[chan], ctx->i32, "");
- }
+ for (chan = 0; chan < count; chan++)
+ address[chan] = ac_to_integer(&ctx->ac, address[chan]);
- /* Adjust the sample index according to FMASK.
- *
- * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
- * which is the identity mapping. Each nibble says which physical sample
- * should be fetched to get that sample.
- *
- * For example, 0x11111100 means there are only 2 samples stored and
- * the second sample covers 3/4 of the pixel. When reading samples 0
- * and 1, return physical sample 0 (determined by the first two 0s
- * in FMASK), otherwise return physical sample 1.
- *
- * The sample index should be adjusted as follows:
- * sample_index = (fmask >> (sample_index * 4)) & 0xF;
- */
if (target == TGSI_TEXTURE_2D_MSAA ||
target == TGSI_TEXTURE_2D_ARRAY_MSAA) {
- struct lp_build_emit_data txf_emit_data = *emit_data;
- LLVMValueRef txf_address[4];
- /* We only need .xy for non-arrays, and .xyz for arrays. */
- unsigned txf_count = target == TGSI_TEXTURE_2D_MSAA ? 2 : 3;
- struct tgsi_full_instruction inst = {};
-
- memcpy(txf_address, address, sizeof(txf_address));
-
- /* Read FMASK using TXF_LZ. */
- inst.Instruction.Opcode = TGSI_OPCODE_TXF_LZ;
- inst.Texture.Texture = target;
- txf_emit_data.inst = &inst;
- txf_emit_data.chan = 0;
- set_tex_fetch_args(ctx, &txf_emit_data,
- target, fmask_ptr, NULL,
- txf_address, txf_count, 0xf);
- build_tex_intrinsic(&tex_action, bld_base, &txf_emit_data);
-
- /* Initialize some constants. */
- LLVMValueRef four = LLVMConstInt(ctx->i32, 4, 0);
- LLVMValueRef F = LLVMConstInt(ctx->i32, 0xF, 0);
-
- /* Apply the formula. */
- LLVMValueRef fmask =
- LLVMBuildExtractElement(gallivm->builder,
- txf_emit_data.output[0],
- ctx->i32_0, "");
-
- unsigned sample_chan = txf_count; /* the sample index is last */
-
- LLVMValueRef sample_index4 =
- LLVMBuildMul(gallivm->builder, address[sample_chan], four, "");
-
- LLVMValueRef shifted_fmask =
- LLVMBuildLShr(gallivm->builder, fmask, sample_index4, "");
-
- LLVMValueRef final_sample =
- LLVMBuildAnd(gallivm->builder, shifted_fmask, F, "");
-
- /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
- * resource descriptor is 0 (invalid),
- */
- LLVMValueRef fmask_desc =
- LLVMBuildBitCast(gallivm->builder, fmask_ptr,
- ctx->v8i32, "");
-
- LLVMValueRef fmask_word1 =
- LLVMBuildExtractElement(gallivm->builder, fmask_desc,
- ctx->i32_1, "");
-
- LLVMValueRef word1_is_nonzero =
- LLVMBuildICmp(gallivm->builder, LLVMIntNE,
- fmask_word1, ctx->i32_0, "");
-
- /* Replace the MSAA sample index. */
- address[sample_chan] =
- LLVMBuildSelect(gallivm->builder, word1_is_nonzero,
- final_sample, address[sample_chan], "");
+ ac_apply_fmask_to_sample(&ctx->ac, fmask_ptr, address,
+ target == TGSI_TEXTURE_2D_ARRAY_MSAA);
}
if (opcode == TGSI_OPCODE_TXF ||
*
* The workaround is to subtract 0.5 from the unnormalized coordinates,
* or (0.5 / size) from the normalized coordinates.
+ *
+ * However, cube textures with 8_8_8_8 data formats require a different
+ * workaround of overriding the num format to USCALED/SSCALED. This would lose
+ * precision in 32-bit data formats, so it needs to be applied dynamically at
+ * runtime. In this case, return an i1 value that indicates whether the
+ * descriptor was overridden (and hence a fixup of the sampler result is needed).
*/
-static void si_lower_gather4_integer(struct si_shader_context *ctx,
- struct ac_image_args *args,
- unsigned target)
+static LLVMValueRef
+si_lower_gather4_integer(struct si_shader_context *ctx,
+ struct ac_image_args *args,
+ unsigned target,
+ enum tgsi_return_type return_type)
{
- LLVMBuilderRef builder = ctx->gallivm.builder;
+ LLVMBuilderRef builder = ctx->ac.builder;
+ LLVMValueRef wa_8888 = NULL;
LLVMValueRef coord = args->addr;
LLVMValueRef half_texel[2];
/* Texture coordinates start after:
unsigned coord_vgpr_index = (int)args->offset + (int)args->compare;
int c;
+ assert(return_type == TGSI_RETURN_TYPE_SINT ||
+ return_type == TGSI_RETURN_TYPE_UINT);
+
+ if (target == TGSI_TEXTURE_CUBE ||
+ target == TGSI_TEXTURE_CUBE_ARRAY) {
+ LLVMValueRef formats;
+ LLVMValueRef data_format;
+ LLVMValueRef wa_formats;
+
+ formats = LLVMBuildExtractElement(builder, args->resource, ctx->i32_1, "");
+
+ data_format = LLVMBuildLShr(builder, formats,
+ LLVMConstInt(ctx->i32, 20, false), "");
+ data_format = LLVMBuildAnd(builder, data_format,
+ LLVMConstInt(ctx->i32, (1u << 6) - 1, false), "");
+ wa_8888 = LLVMBuildICmp(
+ builder, LLVMIntEQ, data_format,
+ LLVMConstInt(ctx->i32, V_008F14_IMG_DATA_FORMAT_8_8_8_8, false),
+ "");
+
+ uint32_t wa_num_format =
+ return_type == TGSI_RETURN_TYPE_UINT ?
+ S_008F14_NUM_FORMAT_GFX6(V_008F14_IMG_NUM_FORMAT_USCALED) :
+ S_008F14_NUM_FORMAT_GFX6(V_008F14_IMG_NUM_FORMAT_SSCALED);
+ wa_formats = LLVMBuildAnd(builder, formats,
+ LLVMConstInt(ctx->i32, C_008F14_NUM_FORMAT_GFX6, false),
+ "");
+ wa_formats = LLVMBuildOr(builder, wa_formats,
+ LLVMConstInt(ctx->i32, wa_num_format, false), "");
+
+ formats = LLVMBuildSelect(builder, wa_8888, wa_formats, formats, "");
+ args->resource = LLVMBuildInsertElement(
+ builder, args->resource, formats, ctx->i32_1, "");
+ }
+
if (target == TGSI_TEXTURE_RECT ||
target == TGSI_TEXTURE_SHADOWRECT) {
+ assert(!wa_8888);
half_texel[0] = half_texel[1] = LLVMConstReal(ctx->f32, -0.5);
} else {
struct tgsi_full_instruction txq_inst = {};
struct lp_build_emit_data txq_emit_data = {};
+ struct lp_build_if_state if_ctx;
+
+ if (wa_8888) {
+ /* Skip the texture size query entirely if we don't need it. */
+ lp_build_if(&if_ctx, &ctx->gallivm, LLVMBuildNot(builder, wa_8888, ""));
+ }
/* Query the texture size. */
txq_inst.Texture.Texture = target;
half_texel[c] = LLVMBuildFMul(builder, half_texel[c],
LLVMConstReal(ctx->f32, -0.5), "");
}
+
+ if (wa_8888) {
+ lp_build_endif(&if_ctx);
+
+ LLVMBasicBlockRef bb[2] = { if_ctx.true_block, if_ctx.entry_block };
+
+ for (c = 0; c < 2; c++) {
+ LLVMValueRef values[2] = { half_texel[c], ctx->ac.f32_0 };
+ half_texel[c] = ac_build_phi(&ctx->ac, ctx->f32, 2,
+ values, bb);
+ }
+ }
}
for (c = 0; c < 2; c++) {
LLVMValueRef index = LLVMConstInt(ctx->i32, coord_vgpr_index + c, 0);
tmp = LLVMBuildExtractElement(builder, coord, index, "");
- tmp = LLVMBuildBitCast(builder, tmp, ctx->f32, "");
+ tmp = ac_to_float(&ctx->ac, tmp);
tmp = LLVMBuildFAdd(builder, tmp, half_texel[c], "");
- tmp = LLVMBuildBitCast(builder, tmp, ctx->i32, "");
+ tmp = ac_to_integer(&ctx->ac, tmp);
coord = LLVMBuildInsertElement(builder, coord, tmp, index, "");
}
args->addr = coord;
+
+ return wa_8888;
+}
+
+/* The second half of the cube texture 8_8_8_8 integer workaround: adjust the
+ * result after the gather operation.
+ */
+static LLVMValueRef
+si_fix_gather4_integer_result(struct si_shader_context *ctx,
+ LLVMValueRef result,
+ enum tgsi_return_type return_type,
+ LLVMValueRef wa)
+{
+ LLVMBuilderRef builder = ctx->ac.builder;
+
+ assert(return_type == TGSI_RETURN_TYPE_SINT ||
+ return_type == TGSI_RETURN_TYPE_UINT);
+
+ for (unsigned chan = 0; chan < 4; ++chan) {
+ LLVMValueRef chanv = LLVMConstInt(ctx->i32, chan, false);
+ LLVMValueRef value;
+ LLVMValueRef wa_value;
+
+ value = LLVMBuildExtractElement(builder, result, chanv, "");
+
+ if (return_type == TGSI_RETURN_TYPE_UINT)
+ wa_value = LLVMBuildFPToUI(builder, value, ctx->i32, "");
+ else
+ wa_value = LLVMBuildFPToSI(builder, value, ctx->i32, "");
+ wa_value = ac_to_float(&ctx->ac, wa_value);
+ value = LLVMBuildSelect(builder, wa, wa_value, value, "");
+
+ result = LLVMBuildInsertElement(builder, result, value, chanv, "");
+ }
+
+ return result;
}
static void build_tex_intrinsic(const struct lp_build_tgsi_action *action,
unsigned target = inst->Texture.Texture;
if (target == TGSI_TEXTURE_BUFFER) {
- emit_data->output[emit_data->chan] =
+ unsigned num_channels =
+ util_last_bit(inst->Dst[0].Register.WriteMask);
+ LLVMValueRef result =
ac_build_buffer_load_format(&ctx->ac,
emit_data->args[0],
emit_data->args[2],
emit_data->args[1],
- true);
+ num_channels, false, true);
+ emit_data->output[emit_data->chan] =
+ ac_build_expand_to_vec4(&ctx->ac, result, num_channels);
return;
}
}
/* The hardware needs special lowering for Gather4 with integer formats. */
- if (ctx->screen->b.chip_class <= VI &&
- opcode == TGSI_OPCODE_TG4) {
- const unsigned src_idx = 2;
+ LLVMValueRef gather4_int_result_workaround = NULL;
- assert(inst->Src[src_idx].Register.File == TGSI_FILE_SAMPLER);
+ if (ctx->screen->info.chip_class <= VI &&
+ opcode == TGSI_OPCODE_TG4) {
assert(inst->Texture.ReturnType != TGSI_RETURN_TYPE_UNKNOWN);
if (inst->Texture.ReturnType == TGSI_RETURN_TYPE_SINT ||
- inst->Texture.ReturnType == TGSI_RETURN_TYPE_UINT)
- si_lower_gather4_integer(ctx, &args, target);
+ inst->Texture.ReturnType == TGSI_RETURN_TYPE_UINT) {
+ gather4_int_result_workaround =
+ si_lower_gather4_integer(ctx, &args, target,
+ inst->Texture.ReturnType);
+ }
}
- emit_data->output[emit_data->chan] =
+ LLVMValueRef result =
ac_build_image_opcode(&ctx->ac, &args);
+
+ if (gather4_int_result_workaround) {
+ result = si_fix_gather4_integer_result(ctx, result,
+ inst->Texture.ReturnType,
+ gather4_int_result_workaround);
+ }
+
+ emit_data->output[emit_data->chan] = result;
}
static void si_llvm_emit_txqs(
struct lp_build_emit_data *emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = &ctx->gallivm;
- LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef res, samples;
LLVMValueRef res_ptr, samp_ptr, fmask_ptr = NULL;
/* Read the samples from the descriptor directly. */
- res = LLVMBuildBitCast(builder, res_ptr, ctx->v8i32, "");
- samples = LLVMBuildExtractElement(
- builder, res,
- LLVMConstInt(ctx->i32, 3, 0), "");
- samples = LLVMBuildLShr(builder, samples,
+ res = LLVMBuildBitCast(ctx->ac.builder, res_ptr, ctx->v8i32, "");
+ samples = LLVMBuildExtractElement(ctx->ac.builder, res,
+ LLVMConstInt(ctx->i32, 3, 0), "");
+ samples = LLVMBuildLShr(ctx->ac.builder, samples,
LLVMConstInt(ctx->i32, 16, 0), "");
- samples = LLVMBuildAnd(builder, samples,
+ samples = LLVMBuildAnd(ctx->ac.builder, samples,
LLVMConstInt(ctx->i32, 0xf, 0), "");
- samples = LLVMBuildShl(builder, ctx->i32_1,
+ samples = LLVMBuildShl(ctx->ac.builder, ctx->i32_1,
samples, "");
emit_data->output[emit_data->chan] = samples;
}
+static void si_llvm_emit_fbfetch(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);
+ struct ac_image_args args = {};
+ LLVMValueRef ptr, image, fmask, addr_vec;
+
+ /* Ignore src0, because KHR_blend_func_extended disallows multiple render
+ * targets.
+ */
+
+ /* Load the image descriptor. */
+ STATIC_ASSERT(SI_PS_IMAGE_COLORBUF0 % 2 == 0);
+ ptr = LLVMGetParam(ctx->main_fn, ctx->param_rw_buffers);
+ ptr = LLVMBuildPointerCast(ctx->ac.builder, ptr,
+ ac_array_in_const32_addr_space(ctx->v8i32), "");
+ image = ac_build_load_to_sgpr(&ctx->ac, ptr,
+ LLVMConstInt(ctx->i32, SI_PS_IMAGE_COLORBUF0 / 2, 0));
+
+ LLVMValueRef addr[4];
+ unsigned chan = 0;
+
+ addr[chan++] = si_unpack_param(ctx, SI_PARAM_POS_FIXED_PT, 0, 16);
+
+ if (!ctx->shader->key.mono.u.ps.fbfetch_is_1D)
+ addr[chan++] = si_unpack_param(ctx, SI_PARAM_POS_FIXED_PT, 16, 16);
+
+ /* Get the current render target layer index. */
+ if (ctx->shader->key.mono.u.ps.fbfetch_layered)
+ addr[chan++] = si_unpack_param(ctx, SI_PARAM_ANCILLARY, 16, 11);
+
+ if (ctx->shader->key.mono.u.ps.fbfetch_msaa)
+ addr[chan++] = si_get_sample_id(ctx);
+
+ while (chan < 4)
+ addr[chan++] = LLVMGetUndef(ctx->i32);
+
+ if (ctx->shader->key.mono.u.ps.fbfetch_msaa) {
+ fmask = ac_build_load_to_sgpr(&ctx->ac, ptr,
+ LLVMConstInt(ctx->i32, SI_PS_IMAGE_COLORBUF0_FMASK / 2, 0));
+
+ ac_apply_fmask_to_sample(&ctx->ac, fmask, addr, false);
+ }
+
+ addr_vec = ac_build_gather_values(&ctx->ac, addr, ARRAY_SIZE(addr));
+
+ args.opcode = ac_image_load;
+ args.resource = image;
+ args.addr = addr_vec;
+ args.dmask = 0xf;
+ args.da = ctx->shader->key.mono.u.ps.fbfetch_layered;
+
+ emit_data->output[emit_data->chan] =
+ ac_build_image_opcode(&ctx->ac, &args);
+}
+
static const struct lp_build_tgsi_action tex_action = {
.fetch_args = tex_fetch_args,
.emit = build_tex_intrinsic,
bld_base->op_actions[TGSI_OPCODE_LODQ] = tex_action;
bld_base->op_actions[TGSI_OPCODE_TXQS].emit = si_llvm_emit_txqs;
+ bld_base->op_actions[TGSI_OPCODE_FBFETCH].emit = si_llvm_emit_fbfetch;
+
bld_base->op_actions[TGSI_OPCODE_LOAD].fetch_args = load_fetch_args;
bld_base->op_actions[TGSI_OPCODE_LOAD].emit = load_emit;
bld_base->op_actions[TGSI_OPCODE_STORE].fetch_args = store_fetch_args;
projects / mesa.git / blobdiff