From: Nicolai Hähnle <nicolai.haeh...@amd.com>
Get most of the churn out of the way before actually loading samplers
via the ABI.
---
src/amd/common/ac_nir_to_llvm.c | 235 ++++++++++++++++++++--------------------
1 file changed, 118 insertions(+), 117 deletions(-)
diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
index 15285d6..5e85408 100644
--- a/src/amd/common/ac_nir_to_llvm.c
+++ b/src/amd/common/ac_nir_to_llvm.c
@@ -1009,21 +1009,21 @@ static LLVMValueRef llvm_extract_elem(struct
ac_llvm_context *ac,
int count = get_llvm_num_components(value);
assert(index < count);
if (count == 1)
return value;
return LLVMBuildExtractElement(ac->builder, value,
LLVMConstInt(ac->i32, index, false), "");
}
-static LLVMValueRef trim_vector(struct nir_to_llvm_context *ctx,
+static LLVMValueRef trim_vector(struct ac_llvm_context *ctx,
LLVMValueRef value, unsigned count)
{
unsigned num_components = get_llvm_num_components(value);
if (count == num_components)
return value;
LLVMValueRef masks[] = {
LLVMConstInt(ctx->i32, 0, false), LLVMConstInt(ctx->i32, 1, false),
LLVMConstInt(ctx->i32, 2, false), LLVMConstInt(ctx->i32, 3, false)};
@@ -1906,41 +1906,41 @@ static void visit_load_const(struct ac_nir_context *ctx,
static LLVMValueRef cast_ptr(struct nir_to_llvm_context *ctx, LLVMValueRef ptr,
LLVMTypeRef type)
{
int addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
return LLVMBuildBitCast(ctx->builder, ptr,
LLVMPointerType(type, addr_space), "");
}
static LLVMValueRef
-get_buffer_size(struct nir_to_llvm_context *ctx, LLVMValueRef descriptor, bool
in_elements)
+get_buffer_size(struct ac_nir_context *ctx, LLVMValueRef descriptor, bool
in_elements)
{
LLVMValueRef size =
- LLVMBuildExtractElement(ctx->builder, descriptor,
- LLVMConstInt(ctx->i32, 2, false), "");
+ LLVMBuildExtractElement(ctx->ac.builder, descriptor,
+ LLVMConstInt(ctx->ac.i32, 2, false),
"");
/* VI only */
- if (ctx->options->chip_class >= VI && in_elements) {
+ if (ctx->abi->chip_class >= VI && in_elements) {
/* 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.
*/
LLVMValueRef stride =
- LLVMBuildExtractElement(ctx->builder, descriptor,
- LLVMConstInt(ctx->i32, 1,
false), "");
- stride = LLVMBuildLShr(ctx->builder, stride,
- LLVMConstInt(ctx->i32, 16, false), "");
- stride = LLVMBuildAnd(ctx->builder, stride,
- LLVMConstInt(ctx->i32, 0x3fff, false),
"");
+ LLVMBuildExtractElement(ctx->ac.builder, descriptor,
+ LLVMConstInt(ctx->ac.i32, 1,
false), "");
+ stride = LLVMBuildLShr(ctx->ac.builder, stride,
+ LLVMConstInt(ctx->ac.i32, 16, false),
"");
+ stride = LLVMBuildAnd(ctx->ac.builder, stride,
+ LLVMConstInt(ctx->ac.i32, 0x3fff, false),
"");
- size = LLVMBuildUDiv(ctx->builder, size, stride, "");
+ size = LLVMBuildUDiv(ctx->ac.builder, size, stride, "");
}
return size;
}
/**
* Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
* intrinsic names).
*/
static void build_int_type_name(
LLVMTypeRef type,
@@ -1948,48 +1948,48 @@ static void build_int_type_name(
{
assert(bufsize >= 6);
if (LLVMGetTypeKind(type) == LLVMVectorTypeKind)
snprintf(buf, bufsize, "v%ui32",
LLVMGetVectorSize(type));
else
strcpy(buf, "i32");
}
-static LLVMValueRef radv_lower_gather4_integer(struct nir_to_llvm_context *ctx,
+static LLVMValueRef radv_lower_gather4_integer(struct ac_llvm_context *ctx,
struct ac_image_args *args,
const nir_tex_instr *instr)
{
enum glsl_base_type stype =
glsl_get_sampler_result_type(instr->texture->var->type);
LLVMValueRef coord = args->addr;
LLVMValueRef half_texel[2];
LLVMValueRef compare_cube_wa;
LLVMValueRef result;
int c;
unsigned coord_vgpr_index = (unsigned)args->offset +
(unsigned)args->compare;
//TODO Rect
{
struct ac_image_args txq_args = { 0 };
txq_args.da = instr->is_array || instr->sampler_dim ==
GLSL_SAMPLER_DIM_CUBE;
txq_args.opcode = ac_image_get_resinfo;
txq_args.dmask = 0xf;
- txq_args.addr = ctx->i32zero;
+ txq_args.addr = ctx->i32_0;
txq_args.resource = args->resource;
- LLVMValueRef size = ac_build_image_opcode(&ctx->ac, &txq_args);
+ LLVMValueRef size = ac_build_image_opcode(ctx, &txq_args);
for (c = 0; c < 2; c++) {
half_texel[c] = LLVMBuildExtractElement(ctx->builder,
size,
LLVMConstInt(ctx->i32, c, false), "");
half_texel[c] = LLVMBuildUIToFP(ctx->builder,
half_texel[c], ctx->f32, "");
- half_texel[c] = ac_build_fdiv(&ctx->ac, ctx->f32one,
half_texel[c]);
+ half_texel[c] = ac_build_fdiv(ctx, ctx->f32_1,
half_texel[c]);
half_texel[c] = LLVMBuildFMul(ctx->builder,
half_texel[c],
LLVMConstReal(ctx->f32,
-0.5), "");
}
}
LLVMValueRef orig_coords = args->addr;
for (c = 0; c < 2; c++) {
LLVMValueRef tmp;
LLVMValueRef index = LLVMConstInt(ctx->i32, coord_vgpr_index +
c, 0);
@@ -2011,50 +2011,50 @@ static LLVMValueRef radv_lower_gather4_integer(struct
nir_to_llvm_context *ctx,
* LLVM ends up dumping SGPRs into VGPRs to deal with the
compare/select,
* and then reads them back. -pro generates two selects,
* one s_cmp for the descriptor rewriting
* one v_cmp for the coordinate and result changes.
*/
if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE) {
LLVMValueRef tmp, tmp2;
/* workaround 8/8/8/8 uint/sint cube gather bug */
/* first detect it then change to a scaled read and f2i */
- tmp = LLVMBuildExtractElement(ctx->builder, args->resource,
ctx->i32one, "");
+ tmp = LLVMBuildExtractElement(ctx->builder, args->resource,
ctx->i32_1, "");
tmp2 = tmp;
/* extract the DATA_FORMAT */
- tmp = ac_build_bfe(&ctx->ac, tmp, LLVMConstInt(ctx->i32, 20,
false),
+ tmp = ac_build_bfe(ctx, tmp, LLVMConstInt(ctx->i32, 20, false),
LLVMConstInt(ctx->i32, 6, false), false);
/* is the DATA_FORMAT == 8_8_8_8 */
compare_cube_wa = LLVMBuildICmp(ctx->builder, LLVMIntEQ, tmp,
LLVMConstInt(ctx->i32, V_008F14_IMG_DATA_FORMAT_8_8_8_8, false), "");
if (stype == GLSL_TYPE_UINT)
/* Create a NUM FORMAT - 0x2 or 0x4 - USCALED or UINT */
tmp = LLVMBuildSelect(ctx->builder, compare_cube_wa,
LLVMConstInt(ctx->i32, 0x8000000, false),
LLVMConstInt(ctx->i32,
0x10000000, false), "");
else
/* Create a NUM FORMAT - 0x3 or 0x5 - SSCALED or SINT */
tmp = LLVMBuildSelect(ctx->builder, compare_cube_wa,
LLVMConstInt(ctx->i32, 0xc000000, false),
LLVMConstInt(ctx->i32,
0x14000000, false), "");
/* replace the NUM FORMAT in the descriptor */
tmp2 = LLVMBuildAnd(ctx->builder, tmp2, LLVMConstInt(ctx->i32,
C_008F14_NUM_FORMAT_GFX6, false), "");
tmp2 = LLVMBuildOr(ctx->builder, tmp2, tmp, "");
- args->resource = LLVMBuildInsertElement(ctx->builder,
args->resource, tmp2, ctx->i32one, "");
+ args->resource = LLVMBuildInsertElement(ctx->builder,
args->resource, tmp2, ctx->i32_1, "");
/* don't modify the coordinates for this case */
coord = LLVMBuildSelect(ctx->builder, compare_cube_wa,
orig_coords, coord, "");
}
args->addr = coord;
- result = ac_build_image_opcode(&ctx->ac, args);
+ result = ac_build_image_opcode(ctx, args);
if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE) {
LLVMValueRef tmp, tmp2;
/* if the cube workaround is in place, f2i the result. */
for (c = 0; c < 4; c++) {
tmp = LLVMBuildExtractElement(ctx->builder, result,
LLVMConstInt(ctx->i32, c, false), "");
if (stype == GLSL_TYPE_UINT)
tmp2 = LLVMBuildFPToUI(ctx->builder, tmp,
ctx->i32, "");
else
@@ -2062,30 +2062,30 @@ static LLVMValueRef radv_lower_gather4_integer(struct
nir_to_llvm_context *ctx,
tmp = LLVMBuildBitCast(ctx->builder, tmp, ctx->i32, "");
tmp2 = LLVMBuildBitCast(ctx->builder, tmp2, ctx->i32,
"");
tmp = LLVMBuildSelect(ctx->builder, compare_cube_wa,
tmp2, tmp, "");
tmp = LLVMBuildBitCast(ctx->builder, tmp, ctx->f32, "");
result = LLVMBuildInsertElement(ctx->builder, result,
tmp, LLVMConstInt(ctx->i32, c, false), "");
}
}
return result;
}
-static LLVMValueRef build_tex_intrinsic(struct nir_to_llvm_context *ctx,
+static LLVMValueRef build_tex_intrinsic(struct ac_nir_context *ctx,
const nir_tex_instr *instr,
bool lod_is_zero,
struct ac_image_args *args)
{
if (instr->sampler_dim == GLSL_SAMPLER_DIM_BUF) {
return ac_build_buffer_load_format(&ctx->ac,
args->resource,
args->addr,
- LLVMConstInt(ctx->i32, 0,
false),
+ LLVMConstInt(ctx->ac.i32, 0,
false),
true);
}
args->opcode = ac_image_sample;
args->compare = instr->is_shadow;
switch (instr->op) {
case nir_texop_txf:
case nir_texop_txf_ms:
case nir_texop_samples_identical:
@@ -2122,21 +2122,21 @@ static LLVMValueRef build_tex_intrinsic(struct
nir_to_llvm_context *ctx,
args->compare = false;
args->offset = false;
break;
default:
break;
}
if (instr->op == nir_texop_tg4) {
enum glsl_base_type stype =
glsl_get_sampler_result_type(instr->texture->var->type);
if (stype == GLSL_TYPE_UINT || stype == GLSL_TYPE_INT) {
- return radv_lower_gather4_integer(ctx, args, instr);
+ return radv_lower_gather4_integer(&ctx->ac, args,
instr);
}
}
return ac_build_image_opcode(&ctx->ac, args);
}
static LLVMValueRef visit_vulkan_resource_index(struct nir_to_llvm_context
*ctx,
nir_intrinsic_instr *instr)
{
LLVMValueRef index = get_src(ctx->nir, instr->src[0]);
unsigned desc_set = nir_intrinsic_desc_set(instr);
@@ -2175,24 +2175,24 @@ static LLVMValueRef visit_load_push_constant(struct
nir_to_llvm_context *ctx,
addr = LLVMConstInt(ctx->i32, nir_intrinsic_base(instr), 0);
addr = LLVMBuildAdd(ctx->builder, addr, get_src(ctx->nir,
instr->src[0]), "");
ptr = ac_build_gep0(&ctx->ac, ctx->push_constants, addr);
ptr = cast_ptr(ctx, ptr, get_def_type(ctx->nir, &instr->dest.ssa));
return LLVMBuildLoad(ctx->builder, ptr, "");
}
-static LLVMValueRef visit_get_buffer_size(struct nir_to_llvm_context *ctx,
+static LLVMValueRef visit_get_buffer_size(struct ac_nir_context *ctx,
const nir_intrinsic_instr *instr)
{
- LLVMValueRef desc = get_src(ctx->nir, instr->src[0]);
+ LLVMValueRef desc = get_src(ctx, instr->src[0]);
return get_buffer_size(ctx, desc, false);
}
static void visit_store_ssbo(struct nir_to_llvm_context *ctx,
nir_intrinsic_instr *instr)
{
const char *store_name;
LLVMValueRef src_data = get_src(ctx->nir, instr->src[0]);
LLVMTypeRef data_type = ctx->f32;
int elem_size_mult = get_elem_bits(&ctx->ac, LLVMTypeOf(src_data)) / 32;
@@ -2206,21 +2206,21 @@ static void visit_store_ssbo(struct nir_to_llvm_context
*ctx,
params[1] = get_src(ctx->nir, instr->src[1]);
params[2] = LLVMConstInt(ctx->i32, 0, false); /* vindex */
params[4] = ctx->i1false; /* glc */
params[5] = ctx->i1false; /* slc */
if (components_32bit > 1)
data_type = LLVMVectorType(ctx->f32, components_32bit);
base_data = to_float(&ctx->ac, src_data);
- base_data = trim_vector(ctx, base_data, instr->num_components);
+ base_data = trim_vector(&ctx->ac, base_data, instr->num_components);
base_data = LLVMBuildBitCast(ctx->builder, base_data,
data_type, "");
base_offset = get_src(ctx->nir, instr->src[2]); /* voffset */
while (writemask) {
int start, count;
LLVMValueRef data;
LLVMValueRef offset;
LLVMValueRef tmp;
u_bit_scan_consecutive_range(&writemask, &start, &count);
@@ -2808,21 +2808,21 @@ load_tes_input(struct nir_to_llvm_context *ctx,
if (instr->variables[0]->var->data.location == VARYING_SLOT_CLIP_DIST0
&&
is_compact && const_index > 3) {
const_index -= 3;
param++;
}
buf_addr = get_tcs_tes_buffer_address_params(ctx, param, const_index,
is_compact, vertex_index,
indir_index);
result = ac_build_buffer_load(&ctx->ac, ctx->hs_ring_tess_offchip,
instr->num_components, NULL,
buf_addr, ctx->oc_lds, is_compact ? (4 *
const_index) : 0, 1, 0, true, false);
- result = trim_vector(ctx, result, instr->num_components);
+ result = trim_vector(&ctx->ac, result, instr->num_components);
result = LLVMBuildBitCast(ctx->builder, result, get_def_type(ctx->nir,
&instr->dest.ssa), "");
return result;
}
static LLVMValueRef
load_gs_input(struct nir_to_llvm_context *ctx,
nir_intrinsic_instr *instr)
{
LLVMValueRef indir_index, vtx_offset;
unsigned const_index;
@@ -3121,21 +3121,21 @@ static int image_type_to_components_count(enum
glsl_sampler_dim dim, bool array)
* 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;
*/
-static LLVMValueRef adjust_sample_index_using_fmask(struct nir_to_llvm_context
*ctx,
+static LLVMValueRef adjust_sample_index_using_fmask(struct ac_llvm_context
*ctx,
LLVMValueRef coord_x,
LLVMValueRef coord_y,
LLVMValueRef coord_z,
LLVMValueRef sample_index,
LLVMValueRef fmask_desc_ptr)
{
LLVMValueRef fmask_load_address[4];
LLVMValueRef res;
fmask_load_address[0] = coord_x;
fmask_load_address[1] = coord_y;
@@ -3143,53 +3143,53 @@ static LLVMValueRef
adjust_sample_index_using_fmask(struct nir_to_llvm_context *
fmask_load_address[2] = coord_z;
fmask_load_address[3] = LLVMGetUndef(ctx->i32);
}
struct ac_image_args args = {0};
args.opcode = ac_image_load;
args.da = coord_z ? true : false;
args.resource = fmask_desc_ptr;
args.dmask = 0xf;
- args.addr = ac_build_gather_values(&ctx->ac, fmask_load_address,
coord_z ? 4 : 2);
+ args.addr = ac_build_gather_values(ctx, fmask_load_address, coord_z ? 4
: 2);
- res = ac_build_image_opcode(&ctx->ac, &args);
+ res = ac_build_image_opcode(ctx, &args);
- res = to_integer(&ctx->ac, res);
+ res = to_integer(ctx, res);
LLVMValueRef four = LLVMConstInt(ctx->i32, 4, false);
LLVMValueRef F = LLVMConstInt(ctx->i32, 0xf, false);
LLVMValueRef fmask = LLVMBuildExtractElement(ctx->builder,
res,
- ctx->i32zero, "");
+ ctx->i32_0, "");
LLVMValueRef sample_index4 =
LLVMBuildMul(ctx->builder, sample_index, four, "");
LLVMValueRef shifted_fmask =
LLVMBuildLShr(ctx->builder, fmask, sample_index4, "");
LLVMValueRef final_sample =
LLVMBuildAnd(ctx->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(ctx->builder, fmask_desc_ptr,
ctx->v8i32, "");
LLVMValueRef fmask_word1 =
LLVMBuildExtractElement(ctx->builder, fmask_desc,
- ctx->i32one, "");
+ ctx->i32_1, "");
LLVMValueRef word1_is_nonzero =
LLVMBuildICmp(ctx->builder, LLVMIntNE,
- fmask_word1, ctx->i32zero, "");
+ fmask_word1, ctx->i32_0, "");
/* Replace the MSAA sample index. */
sample_index =
LLVMBuildSelect(ctx->builder, word1_is_nonzero,
final_sample, sample_index, "");
return sample_index;
}
static LLVMValueRef get_image_coords(struct nir_to_llvm_context *ctx,
const nir_intrinsic_instr *instr)
@@ -3224,21 +3224,21 @@ static LLVMValueRef get_image_coords(struct
nir_to_llvm_context *ctx,
fmask_load_address[0] = LLVMBuildExtractElement(ctx->builder,
src0, masks[0], "");
fmask_load_address[1] = LLVMBuildExtractElement(ctx->builder,
src0, masks[1], "");
if (glsl_sampler_type_is_array(type))
fmask_load_address[2] =
LLVMBuildExtractElement(ctx->builder, src0, masks[2], "");
else
fmask_load_address[2] = NULL;
if (add_frag_pos) {
for (chan = 0; chan < 2; ++chan)
fmask_load_address[chan] =
LLVMBuildAdd(ctx->builder, fmask_load_address[chan],
LLVMBuildFPToUI(ctx->builder, ctx->frag_pos[chan], ctx->i32, ""), "");
}
- sample_index = adjust_sample_index_using_fmask(ctx,
+ sample_index = adjust_sample_index_using_fmask(&ctx->ac,
fmask_load_address[0],
fmask_load_address[1],
fmask_load_address[2],
sample_index,
get_sampler_desc(ctx, instr->variables[0], DESC_FMASK));
}
if (count == 1) {
if (instr->src[0].ssa->num_components)
res = LLVMBuildExtractElement(ctx->builder, src0,
masks[0], "");
else
@@ -3284,21 +3284,21 @@ static LLVMValueRef visit_image_load(struct
nir_to_llvm_context *ctx,
if (glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_BUF) {
params[0] = get_sampler_desc(ctx, instr->variables[0],
DESC_BUFFER);
params[1] = LLVMBuildExtractElement(ctx->builder,
get_src(ctx->nir, instr->src[0]),
LLVMConstInt(ctx->i32, 0,
false), ""); /* vindex */
params[2] = LLVMConstInt(ctx->i32, 0, false); /* voffset */
params[3] = ctx->i1false; /* glc */
params[4] = ctx->i1false; /* slc */
res = ac_build_intrinsic(&ctx->ac,
"llvm.amdgcn.buffer.load.format.v4f32", ctx->v4f32,
params, 5, 0);
- res = trim_vector(ctx, res, instr->dest.ssa.num_components);
+ res = trim_vector(&ctx->ac, res,
instr->dest.ssa.num_components);
res = to_integer(&ctx->ac, res);
} else {
bool is_da = glsl_sampler_type_is_array(type) ||
glsl_get_sampler_dim(type) ==
GLSL_SAMPLER_DIM_CUBE;
LLVMValueRef da = is_da ? ctx->i1true : ctx->i1false;
LLVMValueRef glc = ctx->i1false;
LLVMValueRef slc = ctx->i1false;
params[0] = get_image_coords(ctx, instr);
params[1] = get_sampler_desc(ctx, instr->variables[0],
DESC_IMAGE);
@@ -3452,51 +3452,51 @@ static LLVMValueRef visit_image_atomic(struct
nir_to_llvm_context *ctx,
abort();
}
build_int_type_name(LLVMTypeOf(coords),
coords_type, sizeof(coords_type));
snprintf(intrinsic_name, sizeof(intrinsic_name),
"%s.%s.%s", base_name, atomic_name, coords_type);
return ac_build_intrinsic(&ctx->ac, intrinsic_name, ctx->i32, params,
param_count, 0);
}
-static LLVMValueRef visit_image_size(struct nir_to_llvm_context *ctx,
+static LLVMValueRef visit_image_size(struct ac_nir_context *ctx,
const nir_intrinsic_instr *instr)
{
LLVMValueRef res;
const nir_variable *var = instr->variables[0]->var;
const struct glsl_type *type = instr->variables[0]->var->type;
bool da = glsl_sampler_type_is_array(var->type) ||
glsl_get_sampler_dim(var->type) == GLSL_SAMPLER_DIM_CUBE;
if(instr->variables[0]->deref.child)
type = instr->variables[0]->deref.child->type;
if (glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_BUF)
- return get_buffer_size(ctx, get_sampler_desc(ctx,
instr->variables[0], DESC_BUFFER), true);
+ return get_buffer_size(ctx, get_sampler_desc(ctx->nctx,
instr->variables[0], DESC_BUFFER), true);
struct ac_image_args args = { 0 };
args.da = da;
args.dmask = 0xf;
- args.resource = get_sampler_desc(ctx, instr->variables[0], DESC_IMAGE);
+ args.resource = get_sampler_desc(ctx->nctx, instr->variables[0],
DESC_IMAGE);
args.opcode = ac_image_get_resinfo;
- args.addr = ctx->i32zero;
+ args.addr = ctx->ac.i32_0;
res = ac_build_image_opcode(&ctx->ac, &args);
if (glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_CUBE &&
glsl_sampler_type_is_array(type)) {
- LLVMValueRef two = LLVMConstInt(ctx->i32, 2, false);
- LLVMValueRef six = LLVMConstInt(ctx->i32, 6, false);
- LLVMValueRef z = LLVMBuildExtractElement(ctx->builder, res,
two, "");
- z = LLVMBuildSDiv(ctx->builder, z, six, "");
- res = LLVMBuildInsertElement(ctx->builder, res, z, two, "");
+ LLVMValueRef two = LLVMConstInt(ctx->ac.i32, 2, false);
+ LLVMValueRef six = LLVMConstInt(ctx->ac.i32, 6, false);
+ LLVMValueRef z = LLVMBuildExtractElement(ctx->ac.builder, res,
two, "");
+ z = LLVMBuildSDiv(ctx->ac.builder, z, six, "");
+ res = LLVMBuildInsertElement(ctx->ac.builder, res, z, two, "");
}
return res;
}
#define NOOP_WAITCNT 0xf7f
#define LGKM_CNT 0x07f
#define VM_CNT 0xf70
static void emit_waitcnt(struct nir_to_llvm_context *ctx,
unsigned simm16)
@@ -3951,21 +3951,21 @@ static void visit_intrinsic(struct ac_nir_context *ctx,
case nir_intrinsic_ssbo_atomic_or:
case nir_intrinsic_ssbo_atomic_xor:
case nir_intrinsic_ssbo_atomic_exchange:
case nir_intrinsic_ssbo_atomic_comp_swap:
result = visit_atomic_ssbo(ctx->nctx, instr);
break;
case nir_intrinsic_load_ubo:
result = visit_load_ubo_buffer(ctx, instr);
break;
case nir_intrinsic_get_buffer_size:
- result = visit_get_buffer_size(ctx->nctx, instr);
+ result = visit_get_buffer_size(ctx, instr);
break;
case nir_intrinsic_load_var:
result = visit_load_var(ctx, instr);
break;
case nir_intrinsic_store_var:
visit_store_var(ctx, instr);
break;
case nir_intrinsic_image_load:
result = visit_image_load(ctx->nctx, instr);
break;
@@ -3976,21 +3976,21 @@ static void visit_intrinsic(struct ac_nir_context *ctx,
case nir_intrinsic_image_atomic_min:
case nir_intrinsic_image_atomic_max:
case nir_intrinsic_image_atomic_and:
case nir_intrinsic_image_atomic_or:
case nir_intrinsic_image_atomic_xor:
case nir_intrinsic_image_atomic_exchange:
case nir_intrinsic_image_atomic_comp_swap:
result = visit_image_atomic(ctx->nctx, instr);
break;
case nir_intrinsic_image_size:
- result = visit_image_size(ctx->nctx, instr);
+ result = visit_image_size(ctx, instr);
break;
case nir_intrinsic_discard:
ctx->nctx->shader_info->fs.can_discard = true;
ac_build_intrinsic(&ctx->ac, "llvm.AMDGPU.kilp",
LLVMVoidTypeInContext(ctx->ac.context),
NULL, 0, AC_FUNC_ATTR_LEGACY);
break;
case nir_intrinsic_discard_if:
emit_discard_if(ctx->nctx, instr);
break;
@@ -4117,39 +4117,39 @@ static LLVMValueRef get_sampler_desc(struct
nir_to_llvm_context *ctx,
index = ctx->i32zero;
index = LLVMBuildMul(builder, index, LLVMConstInt(ctx->i32, stride /
type_size, 0), "");
list = ac_build_gep0(&ctx->ac, list, LLVMConstInt(ctx->i32, offset, 0));
list = LLVMBuildPointerCast(builder, list, const_array(type, 0), "");
return ac_build_indexed_load_const(&ctx->ac, list, index);
}
-static void set_tex_fetch_args(struct nir_to_llvm_context *ctx,
+static void set_tex_fetch_args(struct ac_llvm_context *ctx,
struct ac_image_args *args,
const nir_tex_instr *instr,
nir_texop op,
LLVMValueRef res_ptr, LLVMValueRef samp_ptr,
LLVMValueRef *param, unsigned count,
unsigned dmask)
{
unsigned is_rect = 0;
bool da = instr->is_array || instr->sampler_dim ==
GLSL_SAMPLER_DIM_CUBE;
if (op == nir_texop_lod)
da = false;
/* Pad to power of two vector */
while (count < util_next_power_of_two(count))
param[count++] = LLVMGetUndef(ctx->i32);
if (count > 1)
- args->addr = ac_build_gather_values(&ctx->ac, param, count);
+ args->addr = ac_build_gather_values(ctx, param, count);
else
args->addr = param[0];
args->resource = res_ptr;
args->sampler = samp_ptr;
if (instr->sampler_dim == GLSL_SAMPLER_DIM_BUF && op == nir_texop_txf) {
args->addr = param[0];
return;
}
@@ -4203,140 +4203,141 @@ static void tex_fetch_ptrs(struct nir_to_llvm_context
*ctx,
else
*samp_ptr = get_sampler_desc(ctx, instr->texture,
DESC_SAMPLER);
if (instr->sampler_dim < GLSL_SAMPLER_DIM_RECT)
*samp_ptr = sici_fix_sampler_aniso(ctx, *res_ptr,
*samp_ptr);
}
if (fmask_ptr && !instr->sampler && (instr->op == nir_texop_txf_ms ||
instr->op ==
nir_texop_samples_identical))
*fmask_ptr = get_sampler_desc(ctx, instr->texture, DESC_FMASK);
}
-static LLVMValueRef apply_round_slice(struct nir_to_llvm_context *ctx,
+static LLVMValueRef apply_round_slice(struct ac_llvm_context *ctx,
LLVMValueRef coord)
{
- coord = to_float(&ctx->ac, coord);
- coord = ac_build_intrinsic(&ctx->ac, "llvm.rint.f32", ctx->f32, &coord,
1, 0);
- coord = to_integer(&ctx->ac, coord);
+ coord = to_float(ctx, coord);
+ coord = ac_build_intrinsic(ctx, "llvm.rint.f32", ctx->f32, &coord, 1,
0);
+ coord = to_integer(ctx, coord);
return coord;
}
-static void visit_tex(struct nir_to_llvm_context *ctx, nir_tex_instr *instr)
+static void visit_tex(struct ac_nir_context *ctx, nir_tex_instr *instr)
{
LLVMValueRef result = NULL;
struct ac_image_args args = { 0 };
unsigned dmask = 0xf;
LLVMValueRef address[16];
LLVMValueRef coords[5];
LLVMValueRef coord = NULL, lod = NULL, comparator = NULL;
LLVMValueRef bias = NULL, offsets = NULL;
LLVMValueRef res_ptr, samp_ptr, fmask_ptr = NULL, sample_index = NULL;
LLVMValueRef ddx = NULL, ddy = NULL;
LLVMValueRef derivs[6];
unsigned chan, count = 0;
unsigned const_src = 0, num_deriv_comp = 0;
bool lod_is_zero = false;
- tex_fetch_ptrs(ctx, instr, &res_ptr, &samp_ptr, &fmask_ptr);
+
+ tex_fetch_ptrs(ctx->nctx, instr, &res_ptr, &samp_ptr, &fmask_ptr);
for (unsigned i = 0; i < instr->num_srcs; i++) {
switch (instr->src[i].src_type) {
case nir_tex_src_coord:
- coord = get_src(ctx->nir, instr->src[i].src);
+ coord = get_src(ctx, instr->src[i].src);
break;
case nir_tex_src_projector:
break;
case nir_tex_src_comparator:
- comparator = get_src(ctx->nir, instr->src[i].src);
+ comparator = get_src(ctx, instr->src[i].src);
break;
case nir_tex_src_offset:
- offsets = get_src(ctx->nir, instr->src[i].src);
+ offsets = get_src(ctx, instr->src[i].src);
const_src = i;
break;
case nir_tex_src_bias:
- bias = get_src(ctx->nir, instr->src[i].src);
+ bias = get_src(ctx, instr->src[i].src);
break;
case nir_tex_src_lod: {
nir_const_value *val =
nir_src_as_const_value(instr->src[i].src);
if (val && val->i32[0] == 0)
lod_is_zero = true;
- lod = get_src(ctx->nir, instr->src[i].src);
+ lod = get_src(ctx, instr->src[i].src);
break;
}
case nir_tex_src_ms_index:
- sample_index = get_src(ctx->nir, instr->src[i].src);
+ sample_index = get_src(ctx, instr->src[i].src);
break;
case nir_tex_src_ms_mcs:
break;
case nir_tex_src_ddx:
- ddx = get_src(ctx->nir, instr->src[i].src);
+ ddx = get_src(ctx, instr->src[i].src);
num_deriv_comp = instr->src[i].src.ssa->num_components;
break;
case nir_tex_src_ddy:
- ddy = get_src(ctx->nir, instr->src[i].src);
+ ddy = get_src(ctx, instr->src[i].src);
break;
case nir_tex_src_texture_offset:
case nir_tex_src_sampler_offset:
case nir_tex_src_plane:
default:
break;
}
}
if (instr->op == nir_texop_txs && instr->sampler_dim ==
GLSL_SAMPLER_DIM_BUF) {
result = get_buffer_size(ctx, res_ptr, true);
goto write_result;
}
if (instr->op == nir_texop_texture_samples) {
LLVMValueRef res, samples, is_msaa;
- res = LLVMBuildBitCast(ctx->builder, res_ptr, ctx->v8i32, "");
- samples = LLVMBuildExtractElement(ctx->builder, res,
- LLVMConstInt(ctx->i32, 3,
false), "");
- is_msaa = LLVMBuildLShr(ctx->builder, samples,
- LLVMConstInt(ctx->i32, 28, false), "");
- is_msaa = LLVMBuildAnd(ctx->builder, is_msaa,
- LLVMConstInt(ctx->i32, 0xe, false), "");
- is_msaa = LLVMBuildICmp(ctx->builder, LLVMIntEQ, is_msaa,
- LLVMConstInt(ctx->i32, 0xe, false), "");
-
- samples = LLVMBuildLShr(ctx->builder, samples,
- LLVMConstInt(ctx->i32, 16, false), "");
- samples = LLVMBuildAnd(ctx->builder, samples,
- LLVMConstInt(ctx->i32, 0xf, false), "");
- samples = LLVMBuildShl(ctx->builder, ctx->i32one,
+ res = LLVMBuildBitCast(ctx->ac.builder, res_ptr, ctx->ac.v8i32,
"");
+ samples = LLVMBuildExtractElement(ctx->ac.builder, res,
+ LLVMConstInt(ctx->ac.i32, 3,
false), "");
+ is_msaa = LLVMBuildLShr(ctx->ac.builder, samples,
+ LLVMConstInt(ctx->ac.i32, 28, false),
"");
+ is_msaa = LLVMBuildAnd(ctx->ac.builder, is_msaa,
+ LLVMConstInt(ctx->ac.i32, 0xe, false),
"");
+ is_msaa = LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ, is_msaa,
+ LLVMConstInt(ctx->ac.i32, 0xe, false),
"");
+
+ samples = LLVMBuildLShr(ctx->ac.builder, samples,
+ LLVMConstInt(ctx->ac.i32, 16, false),
"");
+ samples = LLVMBuildAnd(ctx->ac.builder, samples,
+ LLVMConstInt(ctx->ac.i32, 0xf, false),
"");
+ samples = LLVMBuildShl(ctx->ac.builder, ctx->ac.i32_1,
samples, "");
- samples = LLVMBuildSelect(ctx->builder, is_msaa, samples,
- ctx->i32one, "");
+ samples = LLVMBuildSelect(ctx->ac.builder, is_msaa, samples,
+ ctx->ac.i32_1, "");
result = samples;
goto write_result;
}
if (coord)
for (chan = 0; chan < instr->coord_components; chan++)
coords[chan] = llvm_extract_elem(&ctx->ac, coord, chan);
if (offsets && instr->op != nir_texop_txf) {
LLVMValueRef offset[3], pack;
for (chan = 0; chan < 3; ++chan)
- offset[chan] = ctx->i32zero;
+ offset[chan] = ctx->ac.i32_0;
args.offset = true;
for (chan = 0; chan < get_llvm_num_components(offsets); chan++)
{
offset[chan] = llvm_extract_elem(&ctx->ac, offsets,
chan);
- offset[chan] = LLVMBuildAnd(ctx->builder, offset[chan],
- LLVMConstInt(ctx->i32,
0x3f, false), "");
+ offset[chan] = LLVMBuildAnd(ctx->ac.builder,
offset[chan],
+ LLVMConstInt(ctx->ac.i32,
0x3f, false), "");
if (chan)
- offset[chan] = LLVMBuildShl(ctx->builder,
offset[chan],
-
LLVMConstInt(ctx->i32, chan * 8, false), "");
+ offset[chan] = LLVMBuildShl(ctx->ac.builder,
offset[chan],
+
LLVMConstInt(ctx->ac.i32, chan * 8, false), "");
}
- pack = LLVMBuildOr(ctx->builder, offset[0], offset[1], "");
- pack = LLVMBuildOr(ctx->builder, pack, offset[2], "");
+ pack = LLVMBuildOr(ctx->ac.builder, offset[0], offset[1], "");
+ pack = LLVMBuildOr(ctx->ac.builder, pack, offset[2], "");
address[count++] = pack;
}
/* pack LOD bias value */
if (instr->op == nir_texop_txb && bias) {
address[count++] = bias;
}
/* Pack depth comparison value */
if (instr->is_shadow && comparator) {
@@ -4360,157 +4361,157 @@ static void visit_tex(struct nir_to_llvm_context
*ctx, nir_tex_instr *instr)
}
for (unsigned i = 0; i < num_deriv_comp; i++) {
derivs[i] = to_float(&ctx->ac,
llvm_extract_elem(&ctx->ac, ddx, i));
derivs[num_deriv_comp + i] = to_float(&ctx->ac,
llvm_extract_elem(&ctx->ac, ddy, i));
}
}
if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE && coord) {
if (instr->is_array && instr->op != nir_texop_lod)
- coords[3] = apply_round_slice(ctx, coords[3]);
+ coords[3] = apply_round_slice(&ctx->ac, coords[3]);
for (chan = 0; chan < instr->coord_components; chan++)
coords[chan] = to_float(&ctx->ac, coords[chan]);
if (instr->coord_components == 3)
- coords[3] = LLVMGetUndef(ctx->f32);
+ coords[3] = LLVMGetUndef(ctx->ac.f32);
ac_prepare_cube_coords(&ctx->ac,
instr->op == nir_texop_txd, instr->is_array,
coords, derivs);
if (num_deriv_comp)
num_deriv_comp--;
}
if (ddx || ddy) {
for (unsigned i = 0; i < num_deriv_comp * 2; i++)
address[count++] = derivs[i];
}
/* Pack texture coordinates */
if (coord) {
address[count++] = coords[0];
if (instr->coord_components > 1) {
if (instr->sampler_dim == GLSL_SAMPLER_DIM_1D &&
instr->is_array && instr->op != nir_texop_txf) {
- coords[1] = apply_round_slice(ctx, coords[1]);
+ coords[1] = apply_round_slice(&ctx->ac,
coords[1]);
}
address[count++] = coords[1];
}
if (instr->coord_components > 2) {
/* This seems like a bit of a hack - but it passes
Vulkan CTS with it */
if (instr->sampler_dim != GLSL_SAMPLER_DIM_3D &&
instr->sampler_dim != GLSL_SAMPLER_DIM_CUBE &&
instr->op != nir_texop_txf) {
- coords[2] = apply_round_slice(ctx, coords[2]);
+ coords[2] = apply_round_slice(&ctx->ac,
coords[2]);
}
address[count++] = coords[2];
}
}
/* Pack LOD */
if (lod && ((instr->op == nir_texop_txl && !lod_is_zero) ||
instr->op == nir_texop_txf)) {
address[count++] = lod;
} else if (instr->op == nir_texop_txf_ms && sample_index) {
address[count++] = sample_index;
} else if(instr->op == nir_texop_txs) {
count = 0;
if (lod)
address[count++] = lod;
else
- address[count++] = ctx->i32zero;
+ address[count++] = ctx->ac.i32_0;
}
for (chan = 0; chan < count; chan++) {
- address[chan] = LLVMBuildBitCast(ctx->builder,
- address[chan], ctx->i32, "");
+ address[chan] = LLVMBuildBitCast(ctx->ac.builder,
+ address[chan], ctx->ac.i32,
"");
}
if (instr->op == nir_texop_samples_identical) {
LLVMValueRef txf_address[4];
struct ac_image_args txf_args = { 0 };
unsigned txf_count = count;
memcpy(txf_address, address, sizeof(txf_address));
if (!instr->is_array)
- txf_address[2] = ctx->i32zero;
- txf_address[3] = ctx->i32zero;
+ txf_address[2] = ctx->ac.i32_0;
+ txf_address[3] = ctx->ac.i32_0;
- set_tex_fetch_args(ctx, &txf_args, instr, nir_texop_txf,
+ set_tex_fetch_args(&ctx->ac, &txf_args, instr, nir_texop_txf,
fmask_ptr, NULL,
txf_address, txf_count, 0xf);
result = build_tex_intrinsic(ctx, instr, false, &txf_args);
- result = LLVMBuildExtractElement(ctx->builder, result,
ctx->i32zero, "");
- result = emit_int_cmp(&ctx->ac, LLVMIntEQ, result,
ctx->i32zero);
+ result = LLVMBuildExtractElement(ctx->ac.builder, result,
ctx->ac.i32_0, "");
+ result = emit_int_cmp(&ctx->ac, LLVMIntEQ, result,
ctx->ac.i32_0);
goto write_result;
}
if (instr->sampler_dim == GLSL_SAMPLER_DIM_MS &&
instr->op != nir_texop_txs) {
unsigned sample_chan = instr->is_array ? 3 : 2;
- address[sample_chan] = adjust_sample_index_using_fmask(ctx,
+ address[sample_chan] = adjust_sample_index_using_fmask(&ctx->ac,
address[0],
address[1],
instr->is_array ? address[2] : NULL,
address[sample_chan],
fmask_ptr);
}
if (offsets && instr->op == nir_texop_txf) {
nir_const_value *const_offset =
nir_src_as_const_value(instr->src[const_src].src);
int num_offsets = instr->src[const_src].src.ssa->num_components;
assert(const_offset);
num_offsets = MIN2(num_offsets, instr->coord_components);
if (num_offsets > 2)
- address[2] = LLVMBuildAdd(ctx->builder,
- address[2],
LLVMConstInt(ctx->i32, const_offset->i32[2], false), "");
+ address[2] = LLVMBuildAdd(ctx->ac.builder,
+ address[2],
LLVMConstInt(ctx->ac.i32, const_offset->i32[2], false), "");
if (num_offsets > 1)
- address[1] = LLVMBuildAdd(ctx->builder,
- address[1],
LLVMConstInt(ctx->i32, const_offset->i32[1], false), "");
- address[0] = LLVMBuildAdd(ctx->builder,
- address[0], LLVMConstInt(ctx->i32,
const_offset->i32[0], false), "");
+ address[1] = LLVMBuildAdd(ctx->ac.builder,
+ address[1],
LLVMConstInt(ctx->ac.i32, const_offset->i32[1], false), "");
+ address[0] = LLVMBuildAdd(ctx->ac.builder,
+ address[0], LLVMConstInt(ctx->ac.i32,
const_offset->i32[0], false), "");
}
/* TODO TG4 support */
if (instr->op == nir_texop_tg4) {
if (instr->is_shadow)
dmask = 1;
else
dmask = 1 << instr->component;
}
- set_tex_fetch_args(ctx, &args, instr, instr->op,
+ set_tex_fetch_args(&ctx->ac, &args, instr, instr->op,
res_ptr, samp_ptr, address, count, dmask);
result = build_tex_intrinsic(ctx, instr, lod_is_zero, &args);
if (instr->op == nir_texop_query_levels)
- result = LLVMBuildExtractElement(ctx->builder, result,
LLVMConstInt(ctx->i32, 3, false), "");
+ result = LLVMBuildExtractElement(ctx->ac.builder, result,
LLVMConstInt(ctx->ac.i32, 3, false), "");
else if (instr->is_shadow && instr->op != nir_texop_txs && instr->op !=
nir_texop_lod && instr->op != nir_texop_tg4)
- result = LLVMBuildExtractElement(ctx->builder, result,
ctx->i32zero, "");
+ result = LLVMBuildExtractElement(ctx->ac.builder, result,
ctx->ac.i32_0, "");
else if (instr->op == nir_texop_txs &&
instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE &&
instr->is_array) {
- LLVMValueRef two = LLVMConstInt(ctx->i32, 2, false);
- LLVMValueRef six = LLVMConstInt(ctx->i32, 6, false);
- LLVMValueRef z = LLVMBuildExtractElement(ctx->builder, result,
two, "");
- z = LLVMBuildSDiv(ctx->builder, z, six, "");
- result = LLVMBuildInsertElement(ctx->builder, result, z, two,
"");
+ LLVMValueRef two = LLVMConstInt(ctx->ac.i32, 2, false);
+ LLVMValueRef six = LLVMConstInt(ctx->ac.i32, 6, false);
+ LLVMValueRef z = LLVMBuildExtractElement(ctx->ac.builder,
result, two, "");
+ z = LLVMBuildSDiv(ctx->ac.builder, z, six, "");
+ result = LLVMBuildInsertElement(ctx->ac.builder, result, z,
two, "");
} else if (instr->dest.ssa.num_components != 4)
- result = trim_vector(ctx, result,
instr->dest.ssa.num_components);
+ result = trim_vector(&ctx->ac, result,
instr->dest.ssa.num_components);
write_result:
if (result) {
assert(instr->dest.is_ssa);
result = to_integer(&ctx->ac, result);
- _mesa_hash_table_insert(ctx->nir->defs, &instr->dest.ssa,
result);
+ _mesa_hash_table_insert(ctx->defs, &instr->dest.ssa, result);
}
}
static void visit_phi(struct ac_nir_context *ctx, nir_phi_instr *instr)
{
LLVMTypeRef type = get_def_type(ctx, &instr->dest.ssa);
LLVMValueRef result = LLVMBuildPhi(ctx->ac.builder, type, "");
_mesa_hash_table_insert(ctx->defs, &instr->dest.ssa, result);
@@ -4585,21 +4586,21 @@ static void visit_block(struct ac_nir_context *ctx,
nir_block *block)
case nir_instr_type_alu:
visit_alu(ctx, nir_instr_as_alu(instr));
break;
case nir_instr_type_load_const:
visit_load_const(ctx, nir_instr_as_load_const(instr));
break;
case nir_instr_type_intrinsic:
visit_intrinsic(ctx, nir_instr_as_intrinsic(instr));
break;
case nir_instr_type_tex:
- visit_tex(ctx->nctx, nir_instr_as_tex(instr));
+ visit_tex(ctx, nir_instr_as_tex(instr));
break;
case nir_instr_type_phi:
visit_phi(ctx, nir_instr_as_phi(instr));
break;
case nir_instr_type_ssa_undef:
visit_ssa_undef(ctx, nir_instr_as_ssa_undef(instr));
break;
case nir_instr_type_jump:
visit_jump(ctx, nir_instr_as_jump(instr));
break;
--
2.9.3
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
