[Mesa-dev] [PATCH 04/28] nir: add support for flushing to zero denorm constants
Connor Abbott
cwabbott0 at gmail.com
Wed Dec 5 17:32:35 UTC 2018
Given that other places call nir_eval_const_opcode(), and they'll be
broken unless they also flush denorms, it's probably a good idea to
move all this into nir_eval_const_opcode() itself.
On Wed, Dec 5, 2018 at 4:56 PM Samuel Iglesias Gonsálvez
<siglesias at igalia.com> wrote:
>
> Signed-off-by: Samuel Iglesias Gonsálvez <siglesias at igalia.com>
> ---
> src/compiler/nir/nir_opt_constant_folding.c | 74 +++++++++++++++++++--
> 1 file changed, 68 insertions(+), 6 deletions(-)
>
> diff --git a/src/compiler/nir/nir_opt_constant_folding.c b/src/compiler/nir/nir_opt_constant_folding.c
> index 1fca530af24..a6df8284e17 100644
> --- a/src/compiler/nir/nir_opt_constant_folding.c
> +++ b/src/compiler/nir/nir_opt_constant_folding.c
> @@ -39,7 +39,7 @@ struct constant_fold_state {
> };
>
> static bool
> -constant_fold_alu_instr(nir_alu_instr *instr, void *mem_ctx)
> +constant_fold_alu_instr(nir_alu_instr *instr, void *mem_ctx, unsigned execution_mode)
> {
> nir_const_value src[NIR_MAX_VEC_COMPONENTS];
>
> @@ -77,12 +77,39 @@ constant_fold_alu_instr(nir_alu_instr *instr, void *mem_ctx)
> switch(load_const->def.bit_size) {
> case 64:
> src[i].u64[j] = load_const->value.u64[instr->src[i].swizzle[j]];
> + if (execution_mode & SHADER_DENORM_FLUSH_TO_ZERO_FP64 &&
> + (nir_op_infos[instr->op].input_types[i] == nir_type_float ||
> + nir_op_infos[instr->op].input_types[i] == nir_type_float64)) {
> + if (src[i].u64[j] < 0x0010000000000000)
> + src[i].u64[j] = 0;
> + if (src[i].u64[j] & 0x8000000000000000 &&
> + !(src[i].u64[j] & 0x7ff0000000000000))
> + src[i].u64[j] = 0x8000000000000000;
> + }
Given that this code is duplicated for inputs and outputs in this
patch, maybe refactor to a shared helper?
> break;
> case 32:
> src[i].u32[j] = load_const->value.u32[instr->src[i].swizzle[j]];
> + if (execution_mode & SHADER_DENORM_FLUSH_TO_ZERO_FP32 &&
> + (nir_op_infos[instr->op].input_types[i] == nir_type_float ||
> + nir_op_infos[instr->op].input_types[i] == nir_type_float32)) {
> + if (src[i].u32[j] < 0x00800000)
> + src[i].u32[j] = 0;
> + if (src[i].u32[j] & 0x80000000 &&
> + !(src[i].u32[j] & 0x7f800000))
> + src[i].u32[j] = 0x80000000;
> + }
> break;
> case 16:
> src[i].u16[j] = load_const->value.u16[instr->src[i].swizzle[j]];
> + if (execution_mode & SHADER_DENORM_FLUSH_TO_ZERO_FP16 &&
> + (nir_op_infos[instr->op].input_types[i] == nir_type_float ||
> + nir_op_infos[instr->op].input_types[i] == nir_type_float16)) {
> + if (src[i].u16[j] < 0x0400)
> + src[i].u16[j] = 0;
> + if (src[i].u16[j] & 0x8000 &&
> + !(src[i].u16[j] & 0x7c00))
> + src[i].u16[j] = 0x8000;
> + }
> break;
> case 8:
> src[i].u8[j] = load_const->value.u8[instr->src[i].swizzle[j]];
> @@ -106,6 +133,40 @@ constant_fold_alu_instr(nir_alu_instr *instr, void *mem_ctx)
> nir_eval_const_opcode(instr->op, instr->dest.dest.ssa.num_components,
> bit_size, src);
>
> + for (unsigned j = 0; j < instr->dest.dest.ssa.num_components; j++) {
> + if (execution_mode & SHADER_DENORM_FLUSH_TO_ZERO_FP64 &&
> + bit_size == 64 &&
> + (nir_op_infos[instr->op].output_type == nir_type_float ||
> + nir_op_infos[instr->op].output_type == nir_type_float64)) {
The bit_size doesn't have to equal the destination bitsize, it's the
bitsize for inputs and outputs which are unsized (e.g. output_type ==
nir_type_float instead of nir_type_float32). This should be
(output_type == nir_type_float && bit_size == 64) || output_type ==
nir_type_float64, and that goes for for the other bitsizes too.
> + if (dest.u64[j] < 0x0010000000000000)
> + dest.u64[j] = 0;
> + if (dest.u64[j] & 0x8000000000000000 &&
> + !(dest.u64[j] & 0x7ff0000000000000))
> + dest.u64[j] = 0x8000000000000000;
> + }
> + if (execution_mode & SHADER_DENORM_FLUSH_TO_ZERO_FP32 &&
> + bit_size == 32 &&
> + (nir_op_infos[instr->op].output_type == nir_type_float ||
> + nir_op_infos[instr->op].output_type == nir_type_float32)) {
> + if (dest.u32[j] < 0x00800000)
> + dest.u32[j] = 0;
> + if (dest.u32[j] & 0x80000000 &&
> + !(dest.u32[j] & 0x7f800000))
> + dest.u32[j] = 0x80000000;
> + }
> +
> + if (execution_mode & SHADER_DENORM_FLUSH_TO_ZERO_FP16 &&
> + bit_size == 16 &&
> + (nir_op_infos[instr->op].output_type == nir_type_float ||
> + nir_op_infos[instr->op].output_type == nir_type_float16)) {
> + if (dest.u16[j] < 0x0400)
> + dest.u16[j] = 0;
> + if (dest.u16[j] & 0x8000 &&
> + !(dest.u16[j] & 0x7c00))
> + dest.u16[j] = 0x8000;
> + }
> + }
> +
> nir_load_const_instr *new_instr =
> nir_load_const_instr_create(mem_ctx,
> instr->dest.dest.ssa.num_components,
> @@ -157,14 +218,14 @@ constant_fold_intrinsic_instr(nir_intrinsic_instr *instr)
> }
>
> static bool
> -constant_fold_block(nir_block *block, void *mem_ctx)
> +constant_fold_block(nir_block *block, void *mem_ctx, unsigned execution_mode)
> {
> bool progress = false;
>
> nir_foreach_instr_safe(instr, block) {
> switch (instr->type) {
> case nir_instr_type_alu:
> - progress |= constant_fold_alu_instr(nir_instr_as_alu(instr), mem_ctx);
> + progress |= constant_fold_alu_instr(nir_instr_as_alu(instr), mem_ctx, execution_mode);
> break;
> case nir_instr_type_intrinsic:
> progress |=
> @@ -180,13 +241,13 @@ constant_fold_block(nir_block *block, void *mem_ctx)
> }
>
> static bool
> -nir_opt_constant_folding_impl(nir_function_impl *impl)
> +nir_opt_constant_folding_impl(nir_function_impl *impl, unsigned execution_mode)
> {
> void *mem_ctx = ralloc_parent(impl);
> bool progress = false;
>
> nir_foreach_block(block, impl) {
> - progress |= constant_fold_block(block, mem_ctx);
> + progress |= constant_fold_block(block, mem_ctx, execution_mode);
> }
>
> if (progress)
> @@ -200,10 +261,11 @@ bool
> nir_opt_constant_folding(nir_shader *shader)
> {
> bool progress = false;
> + unsigned execution_mode = shader->info.shader_float_controls_execution_mode;
>
> nir_foreach_function(function, shader) {
> if (function->impl)
> - progress |= nir_opt_constant_folding_impl(function->impl);
> + progress |= nir_opt_constant_folding_impl(function->impl, execution_mode);
> }
>
> return progress;
> --
> 2.19.1
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev
More information about the mesa-dev
mailing list