Re: [Mesa-dev] [PATCH] radeonsi: eliminate trivial constant VS outputs
Hi Marek
Tested-by: Edmondo Tommasina
I tested the patch with Witcher 2.
Thanks
edmondo
On Tue, Oct 18, 2016 at 6:28 PM, Marek Olšák wrote:
> From: Marek Olšák
>
> These constant value VS PARAM exports:
> - 0,0,0,0
> - 0,0,0,1
> - 1,1,1,0
> - 1,1,1,1
> can be loaded into PS inputs using the DEFAULT_VAL field, and the VS exports
> can be removed from the IR to save export & parameter memory.
>
> After LLVM optimizations, analyze the IR to see which exports are equal to
> the ones listed above (or undef) and remove them if they are.
>
> Targeted use cases:
> - All DX9 eON ports always clear 10 VS outputs to 0.0 even if most of them
> are unused by PS (such as Witcher 2 below).
> - VS output arrays with unused elements that the GLSL compiler can't
> eliminate (such as Batman below).
>
> The shader-db deltas are quite interesting:
> (not from upstream si-report.py, it won't be upstreamed)
>
> PERCENTAGE DELTASShaders PARAM exports (affected only)
> batman_arkham_origins589 -67.17 %
> bioshock-infinite 1769 -0.47 %
> dirt-showdown548 -2.68 %
> dota2 1747 -3.36 %
> f1-2015 776 -4.94 %
> left_4_dead_2 1762 -0.07 %
> metro_2033_redux2670 -0.43 %
> portal 474 -0.22 %
> talos_principle 324 -3.63 %
> warsow 176 -2.20 %
> witcher21040 -73.78 %
>
> All affected 991 -65.37 % ... 9681 -> 3353
>
> Total 26725 -10.82 % ... 58490 -> 52162
> ---
> src/gallium/drivers/radeonsi/si_shader.c| 154
>
> src/gallium/drivers/radeonsi/si_shader.h| 11 ++
> src/gallium/drivers/radeonsi/si_state_shaders.c | 17 ++-
> 3 files changed, 180 insertions(+), 2 deletions(-)
>
> diff --git a/src/gallium/drivers/radeonsi/si_shader.c
> b/src/gallium/drivers/radeonsi/si_shader.c
> index a361418..7fc1df4 100644
> --- a/src/gallium/drivers/radeonsi/si_shader.c
> +++ b/src/gallium/drivers/radeonsi/si_shader.c
> @@ -6593,20 +6593,167 @@ static void si_init_shader_ctx(struct
> si_shader_context *ctx,
> bld_base->op_actions[TGSI_OPCODE_EMIT].emit = si_llvm_emit_vertex;
> bld_base->op_actions[TGSI_OPCODE_ENDPRIM].emit =
> si_llvm_emit_primitive;
> bld_base->op_actions[TGSI_OPCODE_BARRIER].emit = si_llvm_emit_barrier;
>
> bld_base->op_actions[TGSI_OPCODE_MAX].emit =
> build_tgsi_intrinsic_nomem;
> bld_base->op_actions[TGSI_OPCODE_MAX].intr_name = "llvm.maxnum.f32";
> bld_base->op_actions[TGSI_OPCODE_MIN].emit =
> build_tgsi_intrinsic_nomem;
> bld_base->op_actions[TGSI_OPCODE_MIN].intr_name = "llvm.minnum.f32";
> }
>
> +/* Return true if the PARAM export has been eliminated. */
> +static bool si_eliminate_const_output(struct si_shader_context *ctx,
> + LLVMValueRef inst, unsigned offset)
> +{
> + struct si_shader *shader = ctx->shader;
> + unsigned num_outputs = shader->selector->info.num_outputs;
> + double v[4];
> + unsigned i, default_val; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
> +
> + for (i = 0; i < 4; i++) {
> + LLVMBool loses_info;
> + LLVMValueRef p = LLVMGetOperand(inst, 5 + i);
> + if (!LLVMIsConstant(p))
> + return false;
> +
> + /* It's a constant expression. Undef outputs are eliminated
> too. */
> + if (LLVMIsUndef(p))
> + v[i] = 0;
> + else
> + v[i] = LLVMConstRealGetDouble(p, &loses_info);
> +
> + if (v[i] != 0 && v[i] != 1)
> + return false;
> + }
> +
> + /* Only certain combinations of 0 and 1 can be eliminated. */
> + if (v[0] == 0 && v[1] == 0 && v[2] == 0)
> + default_val = v[3] == 0 ? 0 : 1;
> + else if (v[0] == 1 && v[1] == 1 && v[2] == 1)
> + default_val = v[3] == 0 ? 2 : 3;
> + else
> + return false;
> +
> + /* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
> + LLVMInstructionEraseFromParent(inst);
> +
> + /* Change OFFSET to DEFAULT_VAL. */
> + for (i = 0; i < num_outputs; i++) {
> + if (shader->info.vs_output_param_offset[i] == offset) {
> + shader->info.vs_output_param_offset[i] =
> + EXP_PARAM_DEFAULT_VAL_ + default_val;
> + break;
> + }
> + }
> + return true;
> +}
> +
> +struct si_vs_exports {
> + unsigned num;
> + unsigned offset[SI_MAX_VS_OUTPUTS];
> + LLVMValueRef inst[SI_MAX_VS_OUTPUTS];
> +};
> +
> +static void si_eliminate_const_vs_outputs(struct si_shader_context *ctx)
> +{
> + struct si_shader *shader = ctx->shader;
> + struct t
Re: [Mesa-dev] [PATCH] radeonsi: eliminate trivial constant VS outputs
On 18.10.2016 18:28, Marek Olšák wrote:
From: Marek Olšák
These constant value VS PARAM exports:
- 0,0,0,0
- 0,0,0,1
- 1,1,1,0
- 1,1,1,1
can be loaded into PS inputs using the DEFAULT_VAL field, and the VS exports
can be removed from the IR to save export & parameter memory.
After LLVM optimizations, analyze the IR to see which exports are equal to
the ones listed above (or undef) and remove them if they are.
Targeted use cases:
- All DX9 eON ports always clear 10 VS outputs to 0.0 even if most of them
are unused by PS (such as Witcher 2 below).
- VS output arrays with unused elements that the GLSL compiler can't
eliminate (such as Batman below).
The shader-db deltas are quite interesting:
(not from upstream si-report.py, it won't be upstreamed)
PERCENTAGE DELTASShaders PARAM exports (affected only)
batman_arkham_origins589 -67.17 %
bioshock-infinite 1769 -0.47 %
dirt-showdown548 -2.68 %
dota2 1747 -3.36 %
f1-2015 776 -4.94 %
left_4_dead_2 1762 -0.07 %
metro_2033_redux2670 -0.43 %
portal 474 -0.22 %
talos_principle 324 -3.63 %
warsow 176 -2.20 %
witcher21040 -73.78 %
All affected 991 -65.37 % ... 9681 -> 3353
Total 26725 -10.82 % ... 58490 -> 52162
That's a really cool result!
---
src/gallium/drivers/radeonsi/si_shader.c| 154
src/gallium/drivers/radeonsi/si_shader.h| 11 ++
src/gallium/drivers/radeonsi/si_state_shaders.c | 17 ++-
3 files changed, 180 insertions(+), 2 deletions(-)
diff --git a/src/gallium/drivers/radeonsi/si_shader.c
b/src/gallium/drivers/radeonsi/si_shader.c
index a361418..7fc1df4 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -6593,20 +6593,167 @@ static void si_init_shader_ctx(struct
si_shader_context *ctx,
bld_base->op_actions[TGSI_OPCODE_EMIT].emit = si_llvm_emit_vertex;
bld_base->op_actions[TGSI_OPCODE_ENDPRIM].emit = si_llvm_emit_primitive;
bld_base->op_actions[TGSI_OPCODE_BARRIER].emit = si_llvm_emit_barrier;
bld_base->op_actions[TGSI_OPCODE_MAX].emit = build_tgsi_intrinsic_nomem;
bld_base->op_actions[TGSI_OPCODE_MAX].intr_name = "llvm.maxnum.f32";
bld_base->op_actions[TGSI_OPCODE_MIN].emit = build_tgsi_intrinsic_nomem;
bld_base->op_actions[TGSI_OPCODE_MIN].intr_name = "llvm.minnum.f32";
}
+/* Return true if the PARAM export has been eliminated. */
+static bool si_eliminate_const_output(struct si_shader_context *ctx,
+ LLVMValueRef inst, unsigned offset)
+{
+ struct si_shader *shader = ctx->shader;
+ unsigned num_outputs = shader->selector->info.num_outputs;
+ double v[4];
+ unsigned i, default_val; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
+
+ for (i = 0; i < 4; i++) {
+ LLVMBool loses_info;
+ LLVMValueRef p = LLVMGetOperand(inst, 5 + i);
+ if (!LLVMIsConstant(p))
+ return false;
+
+ /* It's a constant expression. Undef outputs are eliminated
too. */
+ if (LLVMIsUndef(p))
+ v[i] = 0;
+ else
+ v[i] = LLVMConstRealGetDouble(p, &loses_info);
+
+ if (v[i] != 0 && v[i] != 1)
+ return false;
+ }
+
+ /* Only certain combinations of 0 and 1 can be eliminated. */
+ if (v[0] == 0 && v[1] == 0 && v[2] == 0)
+ default_val = v[3] == 0 ? 0 : 1;
+ else if (v[0] == 1 && v[1] == 1 && v[2] == 1)
+ default_val = v[3] == 0 ? 2 : 3;
You could slightly improve the handling of undef here by filling two
arrays is_zero[4] and is_one[4] in the loop instead of v[4]. Undef would
set both is_zero and is_one to true. It probably doesn't matter much in
practice, so up to you.
Reviewed-by: Nicolai Hähnle
+ else
+ return false;
+
+ /* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
+ LLVMInstructionEraseFromParent(inst);
+
+ /* Change OFFSET to DEFAULT_VAL. */
+ for (i = 0; i < num_outputs; i++) {
+ if (shader->info.vs_output_param_offset[i] == offset) {
+ shader->info.vs_output_param_offset[i] =
+ EXP_PARAM_DEFAULT_VAL_ + default_val;
+ break;
+ }
+ }
+ return true;
+}
+
+struct si_vs_exports {
+ unsigned num;
+ unsigned offset[SI_MAX_VS_OUTPUTS];
+ LLVMValueRef inst[SI_MAX_VS_OUTPUTS];
+};
+
+static void si_eliminate_const_vs_outputs(struct si_shader_context *ctx)
+{
+ struct si_shader *shader = ctx->shader;
+ struct tgsi_shader_info *info = &shader->sele
[Mesa-dev] [PATCH] radeonsi: eliminate trivial constant VS outputs
From: Marek Olšák
These constant value VS PARAM exports:
- 0,0,0,0
- 0,0,0,1
- 1,1,1,0
- 1,1,1,1
can be loaded into PS inputs using the DEFAULT_VAL field, and the VS exports
can be removed from the IR to save export & parameter memory.
After LLVM optimizations, analyze the IR to see which exports are equal to
the ones listed above (or undef) and remove them if they are.
Targeted use cases:
- All DX9 eON ports always clear 10 VS outputs to 0.0 even if most of them
are unused by PS (such as Witcher 2 below).
- VS output arrays with unused elements that the GLSL compiler can't
eliminate (such as Batman below).
The shader-db deltas are quite interesting:
(not from upstream si-report.py, it won't be upstreamed)
PERCENTAGE DELTASShaders PARAM exports (affected only)
batman_arkham_origins589 -67.17 %
bioshock-infinite 1769 -0.47 %
dirt-showdown548 -2.68 %
dota2 1747 -3.36 %
f1-2015 776 -4.94 %
left_4_dead_2 1762 -0.07 %
metro_2033_redux2670 -0.43 %
portal 474 -0.22 %
talos_principle 324 -3.63 %
warsow 176 -2.20 %
witcher21040 -73.78 %
All affected 991 -65.37 % ... 9681 -> 3353
Total 26725 -10.82 % ... 58490 -> 52162
---
src/gallium/drivers/radeonsi/si_shader.c| 154
src/gallium/drivers/radeonsi/si_shader.h| 11 ++
src/gallium/drivers/radeonsi/si_state_shaders.c | 17 ++-
3 files changed, 180 insertions(+), 2 deletions(-)
diff --git a/src/gallium/drivers/radeonsi/si_shader.c
b/src/gallium/drivers/radeonsi/si_shader.c
index a361418..7fc1df4 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -6593,20 +6593,167 @@ static void si_init_shader_ctx(struct
si_shader_context *ctx,
bld_base->op_actions[TGSI_OPCODE_EMIT].emit = si_llvm_emit_vertex;
bld_base->op_actions[TGSI_OPCODE_ENDPRIM].emit = si_llvm_emit_primitive;
bld_base->op_actions[TGSI_OPCODE_BARRIER].emit = si_llvm_emit_barrier;
bld_base->op_actions[TGSI_OPCODE_MAX].emit = build_tgsi_intrinsic_nomem;
bld_base->op_actions[TGSI_OPCODE_MAX].intr_name = "llvm.maxnum.f32";
bld_base->op_actions[TGSI_OPCODE_MIN].emit = build_tgsi_intrinsic_nomem;
bld_base->op_actions[TGSI_OPCODE_MIN].intr_name = "llvm.minnum.f32";
}
+/* Return true if the PARAM export has been eliminated. */
+static bool si_eliminate_const_output(struct si_shader_context *ctx,
+ LLVMValueRef inst, unsigned offset)
+{
+ struct si_shader *shader = ctx->shader;
+ unsigned num_outputs = shader->selector->info.num_outputs;
+ double v[4];
+ unsigned i, default_val; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
+
+ for (i = 0; i < 4; i++) {
+ LLVMBool loses_info;
+ LLVMValueRef p = LLVMGetOperand(inst, 5 + i);
+ if (!LLVMIsConstant(p))
+ return false;
+
+ /* It's a constant expression. Undef outputs are eliminated
too. */
+ if (LLVMIsUndef(p))
+ v[i] = 0;
+ else
+ v[i] = LLVMConstRealGetDouble(p, &loses_info);
+
+ if (v[i] != 0 && v[i] != 1)
+ return false;
+ }
+
+ /* Only certain combinations of 0 and 1 can be eliminated. */
+ if (v[0] == 0 && v[1] == 0 && v[2] == 0)
+ default_val = v[3] == 0 ? 0 : 1;
+ else if (v[0] == 1 && v[1] == 1 && v[2] == 1)
+ default_val = v[3] == 0 ? 2 : 3;
+ else
+ return false;
+
+ /* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
+ LLVMInstructionEraseFromParent(inst);
+
+ /* Change OFFSET to DEFAULT_VAL. */
+ for (i = 0; i < num_outputs; i++) {
+ if (shader->info.vs_output_param_offset[i] == offset) {
+ shader->info.vs_output_param_offset[i] =
+ EXP_PARAM_DEFAULT_VAL_ + default_val;
+ break;
+ }
+ }
+ return true;
+}
+
+struct si_vs_exports {
+ unsigned num;
+ unsigned offset[SI_MAX_VS_OUTPUTS];
+ LLVMValueRef inst[SI_MAX_VS_OUTPUTS];
+};
+
+static void si_eliminate_const_vs_outputs(struct si_shader_context *ctx)
+{
+ struct si_shader *shader = ctx->shader;
+ struct tgsi_shader_info *info = &shader->selector->info;
+ LLVMBasicBlockRef bb;
+ struct si_vs_exports exports;
+ bool removed_any = false;
+
+ exports.num = 0;
+
+ if ((ctx->type == PIPE_SHADER_VERTEX &&
+(shader->key.vs.as_es || shader->key.vs.as_ls)) ||
+ (ctx->type == PIPE_SHADER_TESS_EVAL && shader->key.tes.as_es))
+ retu
