[Mesa-dev] [PATCH 64/92] ac/nir: add image and write parameter to ac_shader_abi::load_sampler_desc
Nicolai Hähnle
nhaehnle at gmail.com
Mon Jun 26 14:10:43 UTC 2017
From: Nicolai Hähnle <nicolai.haehnle at amd.com>
---
src/amd/common/ac_nir_to_llvm.c | 43 ++++++++++++++++------------
src/amd/common/ac_shader_abi.h | 4 ++-
src/gallium/drivers/radeonsi/si_shader_nir.c | 3 +-
3 files changed, 30 insertions(+), 20 deletions(-)
diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
index c3a885f..26490b3 100644
--- a/src/amd/common/ac_nir_to_llvm.c
+++ b/src/amd/common/ac_nir_to_llvm.c
@@ -183,21 +183,22 @@ struct nir_to_llvm_context {
static inline struct nir_to_llvm_context *
nir_to_llvm_context_from_abi(struct ac_shader_abi *abi)
{
struct nir_to_llvm_context *ctx = NULL;
return container_of(abi, ctx, abi);
}
static LLVMValueRef get_sampler_desc(struct ac_nir_context *ctx,
const nir_deref_var *deref,
- enum ac_descriptor_type desc_type);
+ enum ac_descriptor_type desc_type,
+ bool image, bool write);
static unsigned radeon_llvm_reg_index_soa(unsigned index, unsigned chan)
{
return (index * 4) + chan;
}
static unsigned shader_io_get_unique_index(gl_varying_slot slot)
{
/* handle patch indices separate */
if (slot == VARYING_SLOT_TESS_LEVEL_OUTER)
@@ -3223,21 +3224,21 @@ static LLVMValueRef get_image_coords(struct ac_nir_context *ctx,
fmask_load_address[2] = NULL;
if (add_frag_pos) {
for (chan = 0; chan < 2; ++chan)
fmask_load_address[chan] = LLVMBuildAdd(ctx->ac.builder, fmask_load_address[chan], LLVMBuildFPToUI(ctx->ac.builder, ctx->nctx->frag_pos[chan], ctx->ac.i32, ""), "");
}
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));
+ get_sampler_desc(ctx, instr->variables[0], AC_DESC_FMASK, true, false));
}
if (count == 1) {
if (instr->src[0].ssa->num_components)
res = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[0], "");
else
res = src0;
} else {
int chan;
if (is_ms)
count--;
@@ -3272,40 +3273,40 @@ static LLVMValueRef visit_image_load(struct ac_nir_context *ctx,
const nir_variable *var = instr->variables[0]->var;
const struct glsl_type *type = var->type;
LLVMValueRef i1false = LLVMConstInt(ctx->ac.i1, 0, false);
LLVMValueRef i1true = LLVMConstInt(ctx->ac.i1, 1, false);
if(instr->variables[0]->deref.child)
type = instr->variables[0]->deref.child->type;
type = glsl_without_array(type);
if (glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_BUF) {
- params[0] = get_sampler_desc(ctx, instr->variables[0], AC_DESC_BUFFER);
+ params[0] = get_sampler_desc(ctx, instr->variables[0], AC_DESC_BUFFER, true, false);
params[1] = LLVMBuildExtractElement(ctx->ac.builder, get_src(ctx, instr->src[0]),
ctx->ac.i32_0, ""); /* vindex */
params[2] = ctx->ac.i32_0; /* voffset */
params[3] = i1false; /* glc */
params[4] = i1false; /* slc */
res = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.buffer.load.format.v4f32", ctx->ac.v4f32,
params, 5, 0);
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 ? i1true : i1false;
LLVMValueRef glc = i1false;
LLVMValueRef slc = i1false;
params[0] = get_image_coords(ctx, instr);
- params[1] = get_sampler_desc(ctx, instr->variables[0], AC_DESC_IMAGE);
+ params[1] = get_sampler_desc(ctx, instr->variables[0], AC_DESC_IMAGE, true, false);
params[2] = LLVMConstInt(ctx->ac.i32, 15, false); /* dmask */
if (HAVE_LLVM <= 0x0309) {
params[3] = i1false; /* r128 */
params[4] = da;
params[5] = glc;
params[6] = slc;
} else {
LLVMValueRef lwe = i1false;
params[3] = glc;
params[4] = slc;
@@ -3333,38 +3334,38 @@ static void visit_image_store(struct ac_nir_context *ctx,
const nir_variable *var = instr->variables[0]->var;
const struct glsl_type *type = glsl_without_array(var->type);
LLVMValueRef i1false = LLVMConstInt(ctx->ac.i1, 0, false);
LLVMValueRef i1true = LLVMConstInt(ctx->ac.i1, 1, false);
if (ctx->stage == MESA_SHADER_FRAGMENT)
ctx->nctx->shader_info->fs.writes_memory = true;
if (glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_BUF) {
params[0] = to_float(&ctx->ac, get_src(ctx, instr->src[2])); /* data */
- params[1] = get_sampler_desc(ctx, instr->variables[0], AC_DESC_BUFFER);
+ params[1] = get_sampler_desc(ctx, instr->variables[0], AC_DESC_BUFFER, true, true);
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] = i1false; /* glc */
params[5] = i1false; /* slc */
ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.buffer.store.format.v4f32", ctx->ac.voidt,
params, 6, 0);
} else {
bool is_da = glsl_sampler_type_is_array(type) ||
glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_CUBE;
LLVMValueRef da = is_da ? i1true : i1false;
LLVMValueRef glc = i1false;
LLVMValueRef slc = i1false;
params[0] = 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);
+ params[2] = get_sampler_desc(ctx, instr->variables[0], AC_DESC_IMAGE, true, true);
params[3] = LLVMConstInt(ctx->ac.i32, 15, false); /* dmask */
if (HAVE_LLVM <= 0x0309) {
params[4] = i1false; /* r128 */
params[5] = da;
params[6] = glc;
params[7] = slc;
} else {
LLVMValueRef lwe = i1false;
params[4] = glc;
params[5] = slc;
@@ -3400,32 +3401,34 @@ static LLVMValueRef visit_image_atomic(struct ac_nir_context *ctx,
LLVMValueRef i1true = LLVMConstInt(ctx->ac.i1, 1, false);
if (ctx->stage == MESA_SHADER_FRAGMENT)
ctx->nctx->shader_info->fs.writes_memory = true;
params[param_count++] = get_src(ctx, instr->src[2]);
if (instr->intrinsic == nir_intrinsic_image_atomic_comp_swap)
params[param_count++] = get_src(ctx, instr->src[3]);
if (glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_BUF) {
- params[param_count++] = get_sampler_desc(ctx, instr->variables[0], AC_DESC_BUFFER);
+ params[param_count++] = get_sampler_desc(ctx, instr->variables[0],
+ AC_DESC_BUFFER, true, true);
coords = 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++] = i1false; /* glc */
params[param_count++] = i1false; /* slc */
} else {
bool da = glsl_sampler_type_is_array(type) ||
glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_CUBE;
coords = params[param_count++] = get_image_coords(ctx, instr);
- params[param_count++] = get_sampler_desc(ctx, instr->variables[0], AC_DESC_IMAGE);
+ params[param_count++] = get_sampler_desc(ctx, instr->variables[0],
+ AC_DESC_IMAGE, true, true);
params[param_count++] = i1false; /* r128 */
params[param_count++] = da ? i1true : i1false; /* da */
params[param_count++] = i1false; /* slc */
}
switch (instr->intrinsic) {
case nir_intrinsic_image_atomic_add:
atomic_name = "add";
break;
case nir_intrinsic_image_atomic_min:
@@ -3465,27 +3468,29 @@ static LLVMValueRef visit_image_size(struct ac_nir_context *ctx,
{
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], AC_DESC_BUFFER), true);
+ return get_buffer_size(ctx,
+ get_sampler_desc(ctx, instr->variables[0],
+ AC_DESC_BUFFER, true, false), true);
struct ac_image_args args = { 0 };
args.da = da;
args.dmask = 0xf;
- args.resource = get_sampler_desc(ctx, instr->variables[0], AC_DESC_IMAGE);
+ args.resource = get_sampler_desc(ctx, instr->variables[0], AC_DESC_IMAGE, true, false);
args.opcode = ac_image_get_resinfo;
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->ac.i32, 2, false);
LLVMValueRef six = LLVMConstInt(ctx->ac.i32, 6, false);
LLVMValueRef z = LLVMBuildExtractElement(ctx->ac.builder, res, two, "");
@@ -4039,21 +4044,22 @@ static void visit_intrinsic(struct ac_nir_context *ctx,
if (result) {
_mesa_hash_table_insert(ctx->defs, &instr->dest.ssa, result);
}
}
static LLVMValueRef radv_get_sampler_desc(struct ac_shader_abi *abi,
unsigned descriptor_set,
unsigned base_index,
unsigned constant_index,
LLVMValueRef index,
- enum ac_descriptor_type desc_type)
+ enum ac_descriptor_type desc_type,
+ bool image, bool write)
{
struct nir_to_llvm_context *ctx = nir_to_llvm_context_from_abi(abi);
LLVMValueRef list = ctx->descriptor_sets[descriptor_set];
struct radv_descriptor_set_layout *layout = ctx->options->layout->set[descriptor_set].layout;
struct radv_descriptor_set_binding_layout *binding = layout->binding + base_index;
unsigned offset = binding->offset;
unsigned stride = binding->size;
unsigned type_size;
LLVMBuilderRef builder = ctx->builder;
LLVMTypeRef type;
@@ -4111,21 +4117,22 @@ static LLVMValueRef radv_get_sampler_desc(struct ac_shader_abi *abi,
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 LLVMValueRef get_sampler_desc(struct ac_nir_context *ctx,
const nir_deref_var *deref,
- enum ac_descriptor_type desc_type)
+ enum ac_descriptor_type desc_type,
+ bool image, bool write)
{
LLVMValueRef index = NULL;
unsigned constant_index = 0;
const nir_deref *tail = &deref->deref;
while (tail->child) {
const nir_deref_array *child = nir_deref_as_array(tail->child);
unsigned array_size = glsl_get_aoa_size(tail->child->type);
if (!array_size)
@@ -4147,21 +4154,21 @@ static LLVMValueRef get_sampler_desc(struct ac_nir_context *ctx,
constant_index += child->base_offset * array_size;
tail = &child->deref;
}
return ctx->abi->load_sampler_desc(ctx->abi,
deref->var->data.descriptor_set,
deref->var->data.binding,
constant_index, index,
- desc_type);
+ desc_type, image, write);
}
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)
{
@@ -4220,34 +4227,34 @@ static LLVMValueRef sici_fix_sampler_aniso(struct ac_nir_context *ctx,
return LLVMBuildInsertElement(builder, samp, samp0,
LLVMConstInt(ctx->ac.i32, 0, 0), "");
}
static void tex_fetch_ptrs(struct ac_nir_context *ctx,
nir_tex_instr *instr,
LLVMValueRef *res_ptr, LLVMValueRef *samp_ptr,
LLVMValueRef *fmask_ptr)
{
if (instr->sampler_dim == GLSL_SAMPLER_DIM_BUF)
- *res_ptr = get_sampler_desc(ctx, instr->texture, AC_DESC_BUFFER);
+ *res_ptr = get_sampler_desc(ctx, instr->texture, AC_DESC_BUFFER, false, false);
else
- *res_ptr = get_sampler_desc(ctx, instr->texture, AC_DESC_IMAGE);
+ *res_ptr = get_sampler_desc(ctx, instr->texture, AC_DESC_IMAGE, false, false);
if (samp_ptr) {
if (instr->sampler)
- *samp_ptr = get_sampler_desc(ctx, instr->sampler, AC_DESC_SAMPLER);
+ *samp_ptr = get_sampler_desc(ctx, instr->sampler, AC_DESC_SAMPLER, false, false);
else
- *samp_ptr = get_sampler_desc(ctx, instr->texture, AC_DESC_SAMPLER);
+ *samp_ptr = get_sampler_desc(ctx, instr->texture, AC_DESC_SAMPLER, false, false);
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, AC_DESC_FMASK);
+ *fmask_ptr = get_sampler_desc(ctx, instr->texture, AC_DESC_FMASK, false, false);
}
static LLVMValueRef apply_round_slice(struct ac_llvm_context *ctx,
LLVMValueRef 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;
}
diff --git a/src/amd/common/ac_shader_abi.h b/src/amd/common/ac_shader_abi.h
index c0cdfec..d3c1ee1 100644
--- a/src/amd/common/ac_shader_abi.h
+++ b/src/amd/common/ac_shader_abi.h
@@ -58,20 +58,22 @@ struct ac_shader_abi {
/**
* Load a descriptor associated to a sampler.
*
* \param descriptor_set the descriptor set index (only for Vulkan)
* \param base_index the base index of the sampler variable
* \param constant_index constant part of an array index (or 0, if the
* sampler variable is not an array)
* \param index non-constant part of an array index (may be NULL)
* \param desc_type the type of descriptor to load
+ * \param image whether the descriptor is loaded for an image operation
*/
LLVMValueRef (*load_sampler_desc)(struct ac_shader_abi *abi,
unsigned descriptor_set,
unsigned base_index,
unsigned constant_index,
LLVMValueRef index,
- enum ac_descriptor_type desc_type);
+ enum ac_descriptor_type desc_type,
+ bool image, bool write);
};
#endif /* AC_SHADER_ABI_H */
diff --git a/src/gallium/drivers/radeonsi/si_shader_nir.c b/src/gallium/drivers/radeonsi/si_shader_nir.c
index a278ef2..9ad68f4 100644
--- a/src/gallium/drivers/radeonsi/si_shader_nir.c
+++ b/src/gallium/drivers/radeonsi/si_shader_nir.c
@@ -339,21 +339,22 @@ static void declare_nir_input_fs(struct si_shader_context *ctx,
}
si_llvm_load_input_fs(ctx, *fs_attr_idx, out);
(*fs_attr_idx)++;
}
static LLVMValueRef
si_nir_load_sampler_desc(struct ac_shader_abi *abi,
unsigned descriptor_set, unsigned base_index,
unsigned constant_index, LLVMValueRef dynamic_index,
- enum ac_descriptor_type desc_type)
+ enum ac_descriptor_type desc_type, bool image,
+ bool write)
{
struct si_shader_context *ctx = si_shader_context_from_abi(abi);
LLVMBuilderRef builder = ctx->ac.builder;
LLVMValueRef list = LLVMGetParam(ctx->main_fn, ctx->param_samplers_and_images);
LLVMValueRef index = dynamic_index;
assert(!descriptor_set);
if (!index)
index = ctx->ac.i32_0;
--
2.9.3
More information about the mesa-dev
mailing list