AlexIndustrial/mesa
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

clover/nir: Add an image lowering pass

Browse Source 
Reviewed-by: Dave Airlie <airlied@redhat.com>
Reviewed-by: Karol Herbst <kherbst@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7069>
This commit is contained in:
Jason Ekstrand
2020-08-31 18:09:44 -05:00
committed by Karol Herbst
parent 5e31fad8c9
commit 4f24dee22a
1 changed files with 155 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/gallium/frontends/clover/nir/invocation.cpp
+155
View File
@@ -83,6 +83,158 @@ clover_arg_size_align(const glsl_type *type, unsigned *size, unsigned *align)
}
}
static bool
clover_nir_lower_images(nir_shader *shader)
{
nir_function_impl *impl = nir_shader_get_entrypoint(shader);
ASSERTED int last_loc = -1;
int num_rd_images = 0, num_wr_images = 0, num_samplers = 0;
nir_foreach_uniform_variable(var, shader) {
if (glsl_type_is_image(var->type) || glsl_type_is_sampler(var->type)) {
/* Assume they come in order */
assert(var->data.location > last_loc);
last_loc = var->data.location;
}
/* TODO: Constant samplers */
if (var->type == glsl_bare_sampler_type()) {
var->data.driver_location = num_samplers++;
} else if (glsl_type_is_image(var->type)) {
if (var->data.access & ACCESS_NON_WRITEABLE)
var->data.driver_location = num_rd_images++;
else
var->data.driver_location = num_wr_images++;
} else {
/* CL shouldn't have any sampled images */
assert(!glsl_type_is_sampler(var->type));
}
}
shader->info.num_textures = num_rd_images;
shader->info.textures_used = (1 << num_rd_images) - 1;
shader->info.num_images = num_wr_images;
nir_builder b;
nir_builder_init(&b, impl);
bool progress = false;
nir_foreach_block_reverse(block, impl) {
nir_foreach_instr_reverse_safe(instr, block) {
switch (instr->type) {
case nir_instr_type_deref: {
nir_deref_instr *deref = nir_instr_as_deref(instr);
if (deref->deref_type != nir_deref_type_var)
break;
if (!glsl_type_is_image(deref->type) &&
!glsl_type_is_sampler(deref->type))
break;
b.cursor = nir_instr_remove(&deref->instr);
nir_ssa_def *loc =
nir_imm_intN_t(&b, deref->var->data.driver_location,
deref->dest.ssa.bit_size);
nir_ssa_def_rewrite_uses(&deref->dest.ssa, nir_src_for_ssa(loc));
progress = true;
break;
}
case nir_instr_type_tex: {
nir_tex_instr *tex = nir_instr_as_tex(instr);
unsigned count = 0;
for (unsigned i = 0; i < tex->num_srcs; i++) {
if (tex->src[i].src_type == nir_tex_src_texture_deref ||
tex->src[i].src_type == nir_tex_src_sampler_deref) {
nir_deref_instr *deref = nir_src_as_deref(tex->src[i].src);
if (deref->deref_type == nir_deref_type_var) {
/* In this case, we know the actual variable */
if (tex->src[i].src_type == nir_tex_src_texture_deref)
tex->texture_index = deref->var->data.driver_location;
else
tex->sampler_index = deref->var->data.driver_location;
/* This source gets discarded */
nir_instr_rewrite_src(&tex->instr, &tex->src[i].src,
NIR_SRC_INIT);
continue;
} else {
assert(tex->src[i].src.is_ssa);
b.cursor = nir_before_instr(&tex->instr);
/* Back-ends expect a 32-bit thing, not 64-bit */
nir_ssa_def *offset = nir_u2u32(&b, tex->src[i].src.ssa);
if (tex->src[i].src_type == nir_tex_src_texture_deref)
tex->src[count].src_type = nir_tex_src_texture_offset;
else
tex->src[count].src_type = nir_tex_src_sampler_offset;
nir_instr_rewrite_src(&tex->instr, &tex->src[count].src,
nir_src_for_ssa(offset));
}
} else {
/* If we've removed a source, move this one down */
if (count != i) {
assert(count < i);
tex->src[count].src_type = tex->src[i].src_type;
nir_instr_move_src(&tex->instr, &tex->src[count].src,
&tex->src[i].src);
}
}
count++;
}
tex->num_srcs = count;
progress = true;
break;
}
case nir_instr_type_intrinsic: {
nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
switch (intrin->intrinsic) {
case nir_intrinsic_image_deref_load:
case nir_intrinsic_image_deref_store:
case nir_intrinsic_image_deref_atomic_add:
case nir_intrinsic_image_deref_atomic_imin:
case nir_intrinsic_image_deref_atomic_umin:
case nir_intrinsic_image_deref_atomic_imax:
case nir_intrinsic_image_deref_atomic_umax:
case nir_intrinsic_image_deref_atomic_and:
case nir_intrinsic_image_deref_atomic_or:
case nir_intrinsic_image_deref_atomic_xor:
case nir_intrinsic_image_deref_atomic_exchange:
case nir_intrinsic_image_deref_atomic_comp_swap:
case nir_intrinsic_image_deref_atomic_fadd:
case nir_intrinsic_image_deref_atomic_inc_wrap:
case nir_intrinsic_image_deref_atomic_dec_wrap:
case nir_intrinsic_image_deref_size:
case nir_intrinsic_image_deref_samples: {
assert(intrin->src[0].is_ssa);
b.cursor = nir_before_instr(&intrin->instr);
/* Back-ends expect a 32-bit thing, not 64-bit */
nir_ssa_def *offset = nir_u2u32(&b, intrin->src[0].ssa);
nir_rewrite_image_intrinsic(intrin, offset, false);
progress = true;
break;
}
default:
break;
}
break;
}
default:
break;
}
}
}
if (progress) {
nir_metadata_preserve(impl, nir_metadata_block_index |
nir_metadata_dominance);
} else {
nir_metadata_preserve(impl, nir_metadata_all);
}
return progress;
}
struct clover_lower_nir_state {
std::vector<module::argument> &args;
uint32_t global_dims;
@@ -318,6 +470,9 @@ module clover::nir::spirv_to_nir(const module &mod, const device &dev,
nir_var_function_temp,
glsl_get_cl_type_size_align);
NIR_PASS_V(nir, nir_opt_deref);
NIR_PASS_V(nir, nir_lower_cl_images_to_tex);
NIR_PASS_V(nir, clover_nir_lower_images);
NIR_PASS_V(nir, nir_lower_memcpy);
/* use offsets for kernel inputs (uniform) */