For now I have only enabled this for RADV we can do it
also for radeonsi also but we need to add a CAP for it.
vkpipeline-db results:
Totals from affected shaders:
SGPRS: 4104 -> 3728 (-9.16 %)
VGPRS: 3604 -> 3472 (-3.66 %)
Spilled SGPRs: 0 -> 0 (0.00 %)
Spilled VGPRs: 0 -> 0 (0.00 %)
Private memory VGPRs: 0 -> 0 (0.00 %)
Scratch size: 0 -> 0 (0.00 %) dwords per thread
Code Size: 186868 -> 186740 (-0.07 %) bytes
LDS: 0 -> 0 (0.00 %) blocks
Max Waves: 935 -> 968 (3.53 %)
Wait states: 0 -> 0 (0.00 %)
The VGPR and Max wave changes are in Rise of The Tomb Raider,
there were also SGPR changes in Nier.
---
src/amd/vulkan/radv_pipeline.c | 2 +-
src/compiler/nir/nir.h | 3 +-
src/compiler/nir/nir_linking_helpers.c | 111 +++++++++++++++++++++-
src/mesa/state_tracker/st_glsl_to_nir.cpp | 2 +-
4 files changed, 114 insertions(+), 4 deletions(-)
diff --git a/src/amd/vulkan/radv_pipeline.c b/src/amd/vulkan/radv_pipeline.c
index 8f283365069..99dc54e8ed3 100644
--- a/src/amd/vulkan/radv_pipeline.c
+++ b/src/amd/vulkan/radv_pipeline.c
@@ -1823,7 +1823,7 @@ radv_link_shaders(struct radv_pipeline *pipeline,
nir_shader **shaders)
ordered_shaders[i -
1]);
nir_compact_varyings(ordered_shaders[i],
- ordered_shaders[i - 1], true);
+ ordered_shaders[i - 1], true, true);
if (progress) {
if (nir_lower_global_vars_to_local(ordered_shaders[i]))
{
diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
index 60ea4fbc7ff..e69612c23ae 100644
--- a/src/compiler/nir/nir.h
+++ b/src/compiler/nir/nir.h
@@ -2798,7 +2798,8 @@ bool nir_remove_unused_io_vars(nir_shader *shader, struct
exec_list *var_list,
uint64_t *used_by_other_stage,
uint64_t *used_by_other_stage_patches);
void nir_compact_varyings(nir_shader *producer, nir_shader *consumer,
- bool default_to_smooth_interp);
+ bool default_to_smooth_interp,
+ bool ignore_constant_varyings);
void nir_link_xfb_varyings(nir_shader *producer, nir_shader *consumer);
typedef enum {
diff --git a/src/compiler/nir/nir_linking_helpers.c
b/src/compiler/nir/nir_linking_helpers.c
index 4498d57ac4a..ab29fcc8a1f 100644
--- a/src/compiler/nir/nir_linking_helpers.c
+++ b/src/compiler/nir/nir_linking_helpers.c
@@ -524,6 +524,111 @@ compact_components(nir_shader *producer, nir_shader
*consumer, uint8_t *comps,
&producer->info.patch_outputs_read);
}
+/*
+ * On AMD hardware we can eliminate certain constant outputs between the
+ * vertex and fragment shader. Here we avoid packing varyings with the outputs
+ * we can potentialy eliminate.
+ *
+ * The four constant combination we can eliminate are as follows:
+ *
+ * 0.0, 0.0, 0.0, {0.0,1.0}
+ * 1.0, 1.0, 1.0, {0.0,1.0}
+ *
+ * TODO: We could also try repacking varying constants in order to produce
+ * these combinations.
+ */
+static void
+skip_packing_constant_varyings(nir_shader *producer, nir_shader *consumer,
+ uint8_t *comps)
+{
+ if (consumer->info.stage != MESA_SHADER_FRAGMENT)
+ return;
+
+ if (producer->info.stage != MESA_SHADER_VERTEX &&
+ producer->info.stage != MESA_SHADER_TESS_EVAL)
+ return;
+
+ uint8_t comps_zero[MAX_VARYING] = {0};
+ uint8_t comps_one[MAX_VARYING] = {0};
+ uint32_t skip_locations = 0;
+
+ nir_function_impl *impl = nir_shader_get_entrypoint(producer);
+
+ nir_foreach_block(block, impl) {
+ nir_foreach_instr(instr, block) {
+ if (instr->type != nir_instr_type_intrinsic)
+ continue;
+
+ nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
+
+ if (intr->intrinsic != nir_intrinsic_store_deref)
+ continue;
+
+ nir_variable *var =
+ nir_deref_instr_get_variable(nir_src_as_deref(intr->src[0]));
+
+ if (var->data.location < VARYING_SLOT_VAR0 ||
+ var->data.location - VARYING_SLOT_VAR0 >= MAX_VARYING) {
+ continue;
+ }
+
+ unsigned location = var->data.location - VARYING_SLOT_VAR0;
+ if (skip_locations & (1 << location)) {
+ continue;
+ }
+
+ /* To keep things simple only work with varyings we were able to
+ * previously split.
+ */
+ if (intr->num_components != 1) {
+ skip_locations |= 1 << location;
+ continue;
+ }
+
+ if (glsl_get_base_type(var->type) != GLSL_TYPE_FLOAT) {
+ skip_locations |= 1 << location;
+ continue;
+ }
+
+ if (intr->src[1].ssa->parent_instr->type !=
nir_instr_type_load_const) {
+ skip_locations |= 1 << location;
+ continue;
+ }
+
+ nir_load_const_instr *const_instr =
+ nir_instr_as_load_const(intr->src[1].ssa->parent_instr);
+
+ assert(const_instr->def.bit_size == 32);
+ if (const_instr->value.f32[0] == 0.0) {
+ if (comps_one[location] & 0x7) {
+ skip_locations |= 1 << location;
+ continue;
+ } else {
+ comps_zero[location] |= 1 << var->data.location_frac;
+ }
+ } else if (const_instr->value.f32[0] == 1.0) {
+ if (comps_zero[location] & 0x7) {
+ skip_locations |= 1 << location;
+ continue;
+ } else {
+ comps_one[location] |= 1 << var->data.location_frac;
+ }
+ } else {
+ skip_locations |= 1 << location;
+ }
+ }
+ }
+
+ for (unsigned i = 0; i < MAX_VARYING; i++) {
+ if (skip_locations & 1 << i)
+ continue;
+
+ if (comps_one[i] || comps_zero[i]) {
+ comps[i] = 0xf;
+ }
+ }
+}
+
/* We assume that this has been called more-or-less directly after
* remove_unused_varyings. At this point, all of the varyings that we
* aren't going to be using have been completely removed and the
@@ -536,7 +641,8 @@ compact_components(nir_shader *producer, nir_shader
*consumer, uint8_t *comps,
*/
void
nir_compact_varyings(nir_shader *producer, nir_shader *consumer,
- bool default_to_smooth_interp)
+ bool default_to_smooth_interp,
+ bool ignore_constant_varyings)
{
assert(producer->info.stage != MESA_SHADER_FRAGMENT);
assert(consumer->info.stage != MESA_SHADER_VERTEX);
@@ -545,6 +651,9 @@ nir_compact_varyings(nir_shader *producer, nir_shader
*consumer,
uint8_t interp_type[MAX_VARYINGS_INCL_PATCH] = {0};
uint8_t interp_loc[MAX_VARYINGS_INCL_PATCH] = {0};
+ if (ignore_constant_varyings)
+ skip_packing_constant_varyings(producer, consumer, comps);
+
get_slot_component_masks_and_interp_types(&producer->outputs, comps,
interp_type, interp_loc,
producer->info.stage,
diff --git a/src/mesa/state_tracker/st_glsl_to_nir.cpp
b/src/mesa/state_tracker/st_glsl_to_nir.cpp
index 911284401e0..c6bb1ac1162 100644
--- a/src/mesa/state_tracker/st_glsl_to_nir.cpp
+++ b/src/mesa/state_tracker/st_glsl_to_nir.cpp
@@ -751,7 +751,7 @@ st_link_nir(struct gl_context *ctx,
*/
if (!prev_shader->sh.LinkedTransformFeedback)
nir_compact_varyings(shader_program->_LinkedShaders[prev]->Program->nir,
- nir, ctx->API != API_OPENGL_COMPAT);
+ nir, ctx->API != API_OPENGL_COMPAT, false);
}
prev = i;
}
--
2.17.2
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
