[Mesa-dev] [PATCH 4/5] ac/nir: use ac_build_image_opcode for image intrinsics
Nicolai Hähnle
nhaehnle at gmail.com
Wed Apr 11 11:13:56 UTC 2018
From: Nicolai Hähnle <nicolai.haehnle at amd.com>
So that we'll use the dimension-aware intrinsics in the future.
---
src/amd/common/ac_llvm_build.c | 22 -----
src/amd/common/ac_llvm_build.h | 6 --
src/amd/common/ac_nir_to_llvm.c | 190 +++++++++++++++++-----------------------
3 files changed, 78 insertions(+), 140 deletions(-)
diff --git a/src/amd/common/ac_llvm_build.c b/src/amd/common/ac_llvm_build.c
index 431ec1defb9..b5561533437 100644
--- a/src/amd/common/ac_llvm_build.c
+++ b/src/amd/common/ac_llvm_build.c
@@ -1925,42 +1925,20 @@ LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src0,
one = ctx->f64_1;
}
cmp = LLVMBuildFCmp(ctx->builder, LLVMRealOGT, src0, zero, "");
val = LLVMBuildSelect(ctx->builder, cmp, one, src0, "");
cmp = LLVMBuildFCmp(ctx->builder, LLVMRealOGE, val, zero, "");
val = LLVMBuildSelect(ctx->builder, cmp, val, LLVMConstReal(type, -1.0), "");
return val;
}
-void ac_get_image_intr_name(const char *base_name,
- LLVMTypeRef data_type,
- LLVMTypeRef coords_type,
- LLVMTypeRef rsrc_type,
- char *out_name, unsigned out_len)
-{
- char coords_type_name[8];
-
- ac_build_type_name_for_intr(coords_type, coords_type_name,
- sizeof(coords_type_name));
-
- char data_type_name[8];
- char rsrc_type_name[8];
-
- ac_build_type_name_for_intr(data_type, data_type_name,
- sizeof(data_type_name));
- ac_build_type_name_for_intr(rsrc_type, rsrc_type_name,
- sizeof(rsrc_type_name));
- snprintf(out_name, out_len, "%s.%s.%s.%s", base_name,
- data_type_name, coords_type_name, rsrc_type_name);
-}
-
#define AC_EXP_TARGET (HAVE_LLVM >= 0x0500 ? 0 : 3)
#define AC_EXP_ENABLED_CHANNELS (HAVE_LLVM >= 0x0500 ? 1 : 0)
#define AC_EXP_OUT0 (HAVE_LLVM >= 0x0500 ? 2 : 5)
enum ac_ir_type {
AC_IR_UNDEF,
AC_IR_CONST,
AC_IR_VALUE,
};
diff --git a/src/amd/common/ac_llvm_build.h b/src/amd/common/ac_llvm_build.h
index 6869ac68ab4..d9eb2c72cff 100644
--- a/src/amd/common/ac_llvm_build.h
+++ b/src/amd/common/ac_llvm_build.h
@@ -394,26 +394,20 @@ void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned simm16);
LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0,
unsigned bitsize);
LLVMValueRef ac_build_isign(struct ac_llvm_context *ctx, LLVMValueRef src0,
unsigned bitsize);
LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src0,
unsigned bitsize);
-void ac_get_image_intr_name(const char *base_name,
- LLVMTypeRef data_type,
- LLVMTypeRef coords_type,
- LLVMTypeRef rsrc_type,
- char *out_name, unsigned out_len);
-
void ac_optimize_vs_outputs(struct ac_llvm_context *ac,
LLVMValueRef main_fn,
uint8_t *vs_output_param_offset,
uint32_t num_outputs,
uint8_t *num_param_exports);
void ac_init_exec_full_mask(struct ac_llvm_context *ctx);
void ac_declare_lds_as_pointer(struct ac_llvm_context *ac);
LLVMValueRef ac_lds_load(struct ac_llvm_context *ctx,
LLVMValueRef dw_addr);
diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
index ddd1265c8da..be3a4fd1e66 100644
--- a/src/amd/common/ac_nir_to_llvm.c
+++ b/src/amd/common/ac_nir_to_llvm.c
@@ -1123,55 +1123,39 @@ get_buffer_size(struct ac_nir_context *ctx, LLVMValueRef descriptor, bool in_ele
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->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,
- char *buf, unsigned bufsize)
-{
- assert(bufsize >= 6);
-
- if (LLVMGetTypeKind(type) == LLVMVectorTypeKind)
- snprintf(buf, bufsize, "v%ui32",
- LLVMGetVectorSize(type));
- else
- strcpy(buf, "i32");
-}
-
static LLVMValueRef 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 half_texel[2];
LLVMValueRef compare_cube_wa = NULL;
LLVMValueRef result;
//TODO Rect
{
struct ac_image_args txq_args = { 0 };
txq_args.dim = get_ac_sampler_dim(ctx, instr->sampler_dim, instr->is_array);
txq_args.opcode = ac_image_get_resinfo;
txq_args.dmask = 0xf;
txq_args.lod = ctx->i32_0;
txq_args.resource = args->resource;
+ txq_args.attributes = AC_FUNC_ATTR_READNONE;
LLVMValueRef size = ac_build_image_opcode(ctx, &txq_args);
for (unsigned 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, ctx->f32_1, half_texel[c]);
half_texel[c] = LLVMBuildFMul(ctx->builder, half_texel[c],
LLVMConstReal(ctx->f32, -0.5), "");
}
@@ -1226,20 +1210,22 @@ static LLVMValueRef lower_gather4_integer(struct ac_llvm_context *ctx,
tmp2 = LLVMBuildOr(ctx->builder, tmp2, tmp, "");
args->resource = LLVMBuildInsertElement(ctx->builder, args->resource, tmp2, ctx->i32_1, "");
/* don't modify the coordinates for this case */
for (unsigned c = 0; c < 2; ++c)
args->coords[c] = LLVMBuildSelect(
ctx->builder, compare_cube_wa,
orig_coords[c], args->coords[c], "");
}
+
+ args->attributes = AC_FUNC_ATTR_READNONE;
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 (unsigned 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, "");
@@ -1313,20 +1299,22 @@ static LLVMValueRef build_tex_intrinsic(struct ac_nir_context *ctx,
default:
break;
}
if (instr->op == nir_texop_tg4 && ctx->ac.chip_class <= VI) {
enum glsl_base_type stype = glsl_get_sampler_result_type(instr->texture->var->type);
if (stype == GLSL_TYPE_UINT || stype == GLSL_TYPE_INT) {
return lower_gather4_integer(&ctx->ac, args, instr);
}
}
+
+ args->attributes = AC_FUNC_ATTR_READNONE;
return ac_build_image_opcode(&ctx->ac, args);
}
static LLVMValueRef visit_vulkan_resource_reindex(struct ac_nir_context *ctx,
nir_intrinsic_instr *instr)
{
LLVMValueRef ptr = get_src(ctx, instr->src[0]);
LLVMValueRef index = get_src(ctx, instr->src[1]);
LLVMValueRef result = LLVMBuildGEP(ctx->ac.builder, ptr, &index, 1, "");
@@ -2021,34 +2009,20 @@ static int image_type_to_components_count(enum glsl_sampler_dim dim, bool array)
case GLSL_SAMPLER_DIM_SUBPASS:
return 2;
case GLSL_SAMPLER_DIM_SUBPASS_MS:
return 3;
default:
break;
}
return 0;
}
-static bool
-glsl_is_array_image(const struct glsl_type *type)
-{
- const enum glsl_sampler_dim dim = glsl_get_sampler_dim(type);
-
- if (glsl_sampler_type_is_array(type))
- return true;
-
- return dim == GLSL_SAMPLER_DIM_CUBE ||
- dim == GLSL_SAMPLER_DIM_3D ||
- dim == GLSL_SAMPLER_DIM_SUBPASS ||
- dim == GLSL_SAMPLER_DIM_SUBPASS_MS;
-}
-
/* 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
@@ -2068,20 +2042,21 @@ static LLVMValueRef adjust_sample_index_using_fmask(struct ac_llvm_context *ctx,
args.coords[0] = coord_x;
args.coords[1] = coord_y;
if (coord_z)
args.coords[2] = coord_z;
args.opcode = ac_image_load;
args.dim = coord_z ? ac_image_2darray : ac_image_2d;
args.resource = fmask_desc_ptr;
args.dmask = 0xf;
+ args.attributes = AC_FUNC_ATTR_READNONE;
res = ac_build_image_opcode(ctx, &args);
res = ac_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->i32_0, "");
@@ -2108,32 +2083,31 @@ static LLVMValueRef adjust_sample_index_using_fmask(struct ac_llvm_context *ctx,
LLVMBuildICmp(ctx->builder, LLVMIntNE,
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 ac_nir_context *ctx,
- const nir_intrinsic_instr *instr)
+static void get_image_coords(struct ac_nir_context *ctx,
+ const nir_intrinsic_instr *instr,
+ struct ac_image_args *args)
{
const struct glsl_type *type = glsl_without_array(instr->variables[0]->var->type);
LLVMValueRef src0 = get_src(ctx, instr->src[0]);
- LLVMValueRef coords[4];
LLVMValueRef masks[] = {
LLVMConstInt(ctx->ac.i32, 0, false), LLVMConstInt(ctx->ac.i32, 1, false),
LLVMConstInt(ctx->ac.i32, 2, false), LLVMConstInt(ctx->ac.i32, 3, false),
};
- LLVMValueRef res;
LLVMValueRef sample_index = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[1]), 0);
int count;
enum glsl_sampler_dim dim = glsl_get_sampler_dim(type);
bool is_array = glsl_sampler_type_is_array(type);
bool add_frag_pos = (dim == GLSL_SAMPLER_DIM_SUBPASS ||
dim == GLSL_SAMPLER_DIM_SUBPASS_MS);
bool is_ms = (dim == GLSL_SAMPLER_DIM_MS ||
dim == GLSL_SAMPLER_DIM_SUBPASS_MS);
bool gfx9_1d = ctx->ac.chip_class >= GFX9 && dim == GLSL_SAMPLER_DIM_1D;
@@ -2159,67 +2133,63 @@ static LLVMValueRef get_image_coords(struct ac_nir_context *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], AC_DESC_FMASK, NULL, true, false));
}
if (count == 1 && !gfx9_1d) {
if (instr->src[0].ssa->num_components)
- res = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[0], "");
+ args->coords[0] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[0], "");
else
- res = src0;
+ args->coords[0] = src0;
} else {
int chan;
if (is_ms)
count--;
for (chan = 0; chan < count; ++chan) {
- coords[chan] = ac_llvm_extract_elem(&ctx->ac, src0, chan);
+ args->coords[chan] = ac_llvm_extract_elem(&ctx->ac, src0, chan);
}
if (add_frag_pos) {
- for (chan = 0; chan < 2; ++chan)
- coords[chan] = LLVMBuildAdd(ctx->ac.builder, coords[chan], LLVMBuildFPToUI(ctx->ac.builder, ctx->abi->frag_pos[chan],
+ for (chan = 0; chan < 2; ++chan) {
+ args->coords[chan] = LLVMBuildAdd(
+ ctx->ac.builder, args->coords[chan],
+ LLVMBuildFPToUI(
+ ctx->ac.builder, ctx->abi->frag_pos[chan],
ctx->ac.i32, ""), "");
- coords[2] = ac_to_integer(&ctx->ac, ctx->abi->inputs[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER, 0)]);
+ }
+ args->coords[2] = ac_to_integer(&ctx->ac,
+ ctx->abi->inputs[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER, 0)]);
count++;
}
if (gfx9_1d) {
if (is_array) {
- coords[2] = coords[1];
- coords[1] = ctx->ac.i32_0;
+ args->coords[2] = args->coords[1];
+ args->coords[1] = ctx->ac.i32_0;
} else
- coords[1] = ctx->ac.i32_0;
+ args->coords[1] = ctx->ac.i32_0;
count++;
}
if (is_ms) {
- coords[count] = sample_index;
+ args->coords[count] = sample_index;
count++;
}
-
- if (count == 3) {
- coords[3] = LLVMGetUndef(ctx->ac.i32);
- count = 4;
- }
- res = ac_build_gather_values(&ctx->ac, coords, count);
}
- return res;
}
static LLVMValueRef visit_image_load(struct ac_nir_context *ctx,
const nir_intrinsic_instr *instr)
{
- LLVMValueRef params[7];
LLVMValueRef res;
- char intrinsic_name[64];
const nir_variable *var = instr->variables[0]->var;
const struct glsl_type *type = var->type;
if(instr->variables[0]->deref.child)
type = instr->variables[0]->deref.child->type;
type = glsl_without_array(type);
const enum glsl_sampler_dim dim = glsl_get_sampler_dim(type);
if (dim == GLSL_SAMPLER_DIM_BUF) {
@@ -2233,49 +2203,41 @@ static LLVMValueRef visit_image_load(struct ac_nir_context *ctx,
/* TODO: set "glc" and "can_speculate" when OpenGL needs it. */
res = ac_build_buffer_load_format(&ctx->ac, rsrc, vindex,
ctx->ac.i32_0, num_channels,
false, false);
res = ac_build_expand_to_vec4(&ctx->ac, res, num_channels);
res = ac_trim_vector(&ctx->ac, res, instr->dest.ssa.num_components);
res = ac_to_integer(&ctx->ac, res);
} else {
- LLVMValueRef da = glsl_is_array_image(type) ? ctx->ac.i1true : ctx->ac.i1false;
- LLVMValueRef slc = ctx->ac.i1false;
-
- params[0] = get_image_coords(ctx, instr);
- params[1] = get_sampler_desc(ctx, instr->variables[0], AC_DESC_IMAGE, NULL, true, false);
- params[2] = LLVMConstInt(ctx->ac.i32, 15, false); /* dmask */
- params[3] = (var->data.image._volatile || var->data.image.coherent) ?
- ctx->ac.i1true : ctx->ac.i1false;
- params[4] = slc;
- params[5] = ctx->ac.i1false;
- params[6] = da;
-
- ac_get_image_intr_name("llvm.amdgcn.image.load",
- ctx->ac.v4f32, /* vdata */
- LLVMTypeOf(params[0]), /* coords */
- LLVMTypeOf(params[1]), /* rsrc */
- intrinsic_name, sizeof(intrinsic_name));
-
- res = ac_build_intrinsic(&ctx->ac, intrinsic_name, ctx->ac.v4f32,
- params, 7, AC_FUNC_ATTR_READONLY);
+ struct ac_image_args args = {};
+ args.opcode = ac_image_load;
+ get_image_coords(ctx, instr, &args);
+ args.resource = get_sampler_desc(ctx, instr->variables[0],
+ AC_DESC_IMAGE, NULL, true, false);
+ args.dim = get_ac_image_dim(&ctx->ac, glsl_get_sampler_dim(type),
+ glsl_sampler_type_is_array(type));
+ args.dmask = 15;
+ args.attributes = AC_FUNC_ATTR_READONLY;
+ if (var->data.image._volatile || var->data.image.coherent)
+ args.cache_policy |= ac_glc;
+
+ res = ac_build_image_opcode(&ctx->ac, &args);
}
return ac_to_integer(&ctx->ac, res);
}
static void visit_image_store(struct ac_nir_context *ctx,
nir_intrinsic_instr *instr)
{
LLVMValueRef params[8];
- char intrinsic_name[64];
const nir_variable *var = instr->variables[0]->var;
const struct glsl_type *type = glsl_without_array(var->type);
const enum glsl_sampler_dim dim = glsl_get_sampler_dim(type);
LLVMValueRef glc = ctx->ac.i1false;
bool force_glc = ctx->ac.chip_class == SI;
if (force_glc)
glc = ctx->ac.i1true;
if (dim == GLSL_SAMPLER_DIM_BUF) {
LLVMValueRef rsrc = get_sampler_desc(ctx, instr->variables[0], AC_DESC_BUFFER, NULL, true, true);
@@ -2296,137 +2258,140 @@ static void visit_image_store(struct ac_nir_context *ctx,
params[0] = ac_to_float(&ctx->ac, get_src(ctx, instr->src[2])); /* data */
params[1] = rsrc;
params[2] = LLVMBuildExtractElement(ctx->ac.builder, get_src(ctx, instr->src[0]),
ctx->ac.i32_0, ""); /* vindex */
params[3] = ctx->ac.i32_0; /* voffset */
params[4] = glc; /* glc */
params[5] = ctx->ac.i1false; /* slc */
ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.buffer.store.format.v4f32", ctx->ac.voidt,
params, 6, 0);
} else {
- LLVMValueRef da = glsl_is_array_image(type) ? ctx->ac.i1true : ctx->ac.i1false;
- LLVMValueRef slc = ctx->ac.i1false;
-
- params[0] = ac_to_float(&ctx->ac, get_src(ctx, instr->src[2]));
- params[1] = get_image_coords(ctx, instr); /* coords */
- params[2] = get_sampler_desc(ctx, instr->variables[0], AC_DESC_IMAGE, NULL, true, true);
- params[3] = LLVMConstInt(ctx->ac.i32, 15, false); /* dmask */
- params[4] = (force_glc || var->data.image._volatile || var->data.image.coherent) ?
- ctx->ac.i1true : ctx->ac.i1false;
- params[5] = slc;
- params[6] = ctx->ac.i1false;
- params[7] = da;
+ struct ac_image_args args = {};
+ args.opcode = ac_image_store;
+ args.data[0] = ac_to_float(&ctx->ac, get_src(ctx, instr->src[2]));
+ get_image_coords(ctx, instr, &args);
+ args.resource = get_sampler_desc(ctx, instr->variables[0],
+ AC_DESC_IMAGE, NULL, true, false);
+ args.dim = get_ac_image_dim(&ctx->ac, glsl_get_sampler_dim(type),
+ glsl_sampler_type_is_array(type));
+ args.dmask = 15;
+ if (force_glc || var->data.image._volatile || var->data.image.coherent)
+ args.cache_policy |= ac_glc;
- ac_get_image_intr_name("llvm.amdgcn.image.store",
- LLVMTypeOf(params[0]), /* vdata */
- LLVMTypeOf(params[1]), /* coords */
- LLVMTypeOf(params[2]), /* rsrc */
- intrinsic_name, sizeof(intrinsic_name));
-
- ac_build_intrinsic(&ctx->ac, intrinsic_name, ctx->ac.voidt,
- params, 8, 0);
+ ac_build_image_opcode(&ctx->ac, &args);
}
}
static LLVMValueRef visit_image_atomic(struct ac_nir_context *ctx,
const nir_intrinsic_instr *instr)
{
LLVMValueRef params[7];
int param_count = 0;
const nir_variable *var = instr->variables[0]->var;
+ bool cmpswap = instr->intrinsic == nir_intrinsic_image_var_atomic_comp_swap;
const char *atomic_name;
char intrinsic_name[41];
+ enum ac_atomic_op atomic_subop;
const struct glsl_type *type = glsl_without_array(var->type);
MAYBE_UNUSED int length;
bool is_unsigned = glsl_get_sampler_result_type(type) == GLSL_TYPE_UINT;
switch (instr->intrinsic) {
case nir_intrinsic_image_var_atomic_add:
atomic_name = "add";
+ atomic_subop = ac_atomic_add;
break;
case nir_intrinsic_image_var_atomic_min:
atomic_name = is_unsigned ? "umin" : "smin";
+ atomic_subop = is_unsigned ? ac_atomic_umin : ac_atomic_smin;
break;
case nir_intrinsic_image_var_atomic_max:
atomic_name = is_unsigned ? "umax" : "smax";
+ atomic_subop = is_unsigned ? ac_atomic_umax : ac_atomic_smax;
break;
case nir_intrinsic_image_var_atomic_and:
atomic_name = "and";
+ atomic_subop = ac_atomic_and;
break;
case nir_intrinsic_image_var_atomic_or:
atomic_name = "or";
+ atomic_subop = ac_atomic_or;
break;
case nir_intrinsic_image_var_atomic_xor:
atomic_name = "xor";
+ atomic_subop = ac_atomic_xor;
break;
case nir_intrinsic_image_var_atomic_exchange:
atomic_name = "swap";
+ atomic_subop = ac_atomic_swap;
break;
case nir_intrinsic_image_var_atomic_comp_swap:
atomic_name = "cmpswap";
+ atomic_subop = 0; /* not used */
break;
default:
abort();
}
- if (instr->intrinsic == nir_intrinsic_image_var_atomic_comp_swap)
+ if (cmpswap)
params[param_count++] = get_src(ctx, instr->src[3]);
params[param_count++] = get_src(ctx, instr->src[2]);
if (glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_BUF) {
params[param_count++] = get_sampler_desc(ctx, instr->variables[0], AC_DESC_BUFFER,
NULL, true, true);
params[param_count++] = LLVMBuildExtractElement(ctx->ac.builder, get_src(ctx, instr->src[0]),
ctx->ac.i32_0, ""); /* vindex */
params[param_count++] = ctx->ac.i32_0; /* voffset */
params[param_count++] = ctx->ac.i1false; /* slc */
length = snprintf(intrinsic_name, sizeof(intrinsic_name),
"llvm.amdgcn.buffer.atomic.%s", atomic_name);
- } else {
- char coords_type[8];
-
- LLVMValueRef coords = params[param_count++] = get_image_coords(ctx, instr);
- params[param_count++] = get_sampler_desc(ctx, instr->variables[0], AC_DESC_IMAGE,
- NULL, true, true);
- params[param_count++] = ctx->ac.i1false; /* r128 */
- params[param_count++] = glsl_is_array_image(type) ? ctx->ac.i1true : ctx->ac.i1false; /* da */
- params[param_count++] = ctx->ac.i1false; /* slc */
- build_int_type_name(LLVMTypeOf(coords),
- coords_type, sizeof(coords_type));
+ assert(length < sizeof(intrinsic_name));
+ return ac_build_intrinsic(&ctx->ac, intrinsic_name, ctx->ac.i32,
+ params, param_count, 0);
+ } else {
+ struct ac_image_args args = {};
+ args.opcode = cmpswap ? ac_image_atomic_cmpswap : ac_image_atomic;
+ args.atomic = atomic_subop;
+ args.data[0] = params[0];
+ if (cmpswap)
+ args.data[1] = params[1];
+ get_image_coords(ctx, instr, &args);
+ args.resource = get_sampler_desc(ctx, instr->variables[0],
+ AC_DESC_IMAGE, NULL, true, false);
+ args.dim = get_ac_image_dim(&ctx->ac, glsl_get_sampler_dim(type),
+ glsl_sampler_type_is_array(type));
- length = snprintf(intrinsic_name, sizeof(intrinsic_name),
- "llvm.amdgcn.image.atomic.%s.%s", atomic_name, coords_type);
+ return ac_build_image_opcode(&ctx->ac, &args);
}
-
- assert(length < sizeof(intrinsic_name));
- return ac_build_intrinsic(&ctx->ac, intrinsic_name, ctx->ac.i32, params, param_count, 0);
}
static LLVMValueRef visit_image_samples(struct ac_nir_context *ctx,
const nir_intrinsic_instr *instr)
{
const nir_variable *var = instr->variables[0]->var;
const struct glsl_type *type = glsl_without_array(var->type);
struct ac_image_args args = { 0 };
args.dim = get_ac_sampler_dim(&ctx->ac, glsl_get_sampler_dim(type),
glsl_sampler_type_is_array(type));
args.dmask = 0xf;
args.resource = get_sampler_desc(ctx, instr->variables[0],
AC_DESC_IMAGE, NULL, true, false);
args.opcode = ac_image_get_resinfo;
args.lod = ctx->ac.i32_0;
+ args.attributes = AC_FUNC_ATTR_READNONE;
return ac_build_image_opcode(&ctx->ac, &args);
}
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 = glsl_without_array(var->type);
@@ -2437,20 +2402,21 @@ static LLVMValueRef visit_image_size(struct ac_nir_context *ctx,
AC_DESC_BUFFER, NULL, true, false), true);
struct ac_image_args args = { 0 };
args.dim = get_ac_image_dim(&ctx->ac, glsl_get_sampler_dim(type),
glsl_sampler_type_is_array(type));
args.dmask = 0xf;
args.resource = get_sampler_desc(ctx, instr->variables[0], AC_DESC_IMAGE, NULL, true, false);
args.opcode = ac_image_get_resinfo;
args.lod = ctx->ac.i32_0;
+ args.attributes = AC_FUNC_ATTR_READNONE;
res = ac_build_image_opcode(&ctx->ac, &args);
LLVMValueRef two = LLVMConstInt(ctx->ac.i32, 2, false);
if (glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_CUBE &&
glsl_sampler_type_is_array(type)) {
LLVMValueRef six = LLVMConstInt(ctx->ac.i32, 6, false);
LLVMValueRef z = LLVMBuildExtractElement(ctx->ac.builder, res, two, "");
z = LLVMBuildSDiv(ctx->ac.builder, z, six, "");
--
2.14.1
More information about the mesa-dev
mailing list