nir: replace lower_sincos with algebraic opt

This version has less ops for the same precision.

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Vasily Khoruzhick <anarsoul@gmail.com>
Acked-by: Matt Turner <mattst88@gmail.com>

src/compiler/Makefile.sources

@@ -273,7 +273,6 @@ NIR_FILES = \
 	nir/nir_lower_regs_to_ssa.c \
 	nir/nir_lower_returns.c \
 	nir/nir_lower_scratch.c \
-	nir/nir_lower_sincos.c \
 	nir/nir_lower_subgroups.c \
 	nir/nir_lower_system_values.c \
 	nir/nir_lower_tex.c \

src/compiler/nir/meson.build

@@ -152,7 +152,6 @@ files_libnir = files(
   'nir_lower_returns.c',
   'nir_lower_scratch.c',
   'nir_lower_subgroups.c',
-  'nir_lower_sincos.c',
   'nir_lower_system_values.c',
   'nir_lower_tex.c',
   'nir_lower_to_source_mods.c',

src/compiler/nir/nir.h

@@ -2505,6 +2505,7 @@ typedef struct nir_shader_compiler_options {
    bool lower_fpow;
    bool lower_fsat;
    bool lower_fsqrt;
+   bool lower_sincos;
    bool lower_fmod;
    /** Lowers ibitfield_extract/ubitfield_extract to ibfe/ubfe. */
    bool lower_bitfield_extract;
@@ -3976,8 +3977,6 @@ uint64_t nir_get_single_slot_attribs_mask(uint64_t attribs, uint64_t dual_slot);
 nir_intrinsic_op nir_intrinsic_from_system_value(gl_system_value val);
 gl_system_value nir_system_value_from_intrinsic(nir_intrinsic_op intrin);
 
-bool nir_lower_sincos(nir_shader *shader);
-
 static inline bool
 nir_variable_is_in_ubo(const nir_variable *var)
 {

src/compiler/nir/nir_lower_sincos.c

@@ -1,138 +0,0 @@
-/*
- * Copyright (c) 2014 Scott Mansell
- * Copyright © 2014 Broadcom
- * Copyright (c) 2019 Vasily Khoruzhick <anarsoul@gmail.com>
- *
- * Permission is hereby granted, free of charge, to any person obtaining a
- * copy of this software and associated documentation files (the "Software"),
- * to deal in the Software without restriction, including without limitation
- * the rights to use, copy, modify, merge, publish, distribute, sublicense,
- * and/or sell copies of the Software, and to permit persons to whom the
- * Software is furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice (including the next
- * paragraph) shall be included in all copies or substantial portions of the
- * Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
- * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
- * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
- * IN THE SOFTWARE.
- */
-
-#include "util/u_math.h"
-#include "nir.h"
-#include "nir_builder.h"
-
-static nir_ssa_def *
-shrink_input(nir_builder *b, nir_ssa_def *x)
-{
-   nir_ssa_def *scaled_x = nir_fmul_imm(b, x, 1.0 / (M_PI * 2));
-
-   nir_ssa_def *xfrac = nir_ffract(b, scaled_x);
-   /* Map [0.5, 1] to [-0.5, 0] */
-   nir_ssa_def *xfrac_0_5_1 = nir_fadd_imm(b, xfrac, -1.0);
-   /* Map [-1, -0.5] to [0, 0.5] */
-   nir_ssa_def *xfrac_n1_n0_5 = nir_fadd_imm(b, xfrac, 1.0);
-
-   nir_ssa_def *geq_0_5 = nir_build_alu(b, nir_op_fge, xfrac, nir_imm_float(b, 0.5), NULL, NULL);
-   nir_ssa_def *less_m0_5 = nir_build_alu(b, nir_op_flt, xfrac, nir_imm_float(b, -0.5), NULL, NULL);
-
-   nir_ssa_def *sel1 = nir_build_alu(b, nir_op_bcsel, geq_0_5, xfrac_0_5_1, xfrac, NULL );
-   nir_ssa_def *sel2 = nir_build_alu(b, nir_op_bcsel, less_m0_5, xfrac_n1_n0_5, sel1, NULL);
-
-   return sel2;
-}
-
-static nir_ssa_def *
-lower_sincos(nir_builder *b, nir_ssa_def *src, bool do_cos)
-{
-   /* Fast sin/cos implementation, see
-    * https://web.archive.org/web/20180105155939/http://forum.devmaster.net/t/fast-and-accurate-sine-cosine/9648
-    */
-   const float B = 8.0; //(M_PI * 2.0) * (4.0 / M_PI);
-   const float C = -16.0; //(M_PI * 2.0) * (M_PI * 2.0) * (-4.0 / (M_PI * M_PI));
-   const float P = 0.225;
-
-   if (do_cos)
-      src = nir_fadd_imm(b, src, M_PI / 2.0);
-
-   nir_ssa_def *x = shrink_input(b, src);
-
-   nir_ssa_def *bx = nir_fmul_imm(b, x, B);
-   nir_ssa_def *cx = nir_fmul_imm(b, x, C);
-   nir_ssa_def *absx = nir_fabs(b, x);
-   nir_ssa_def *cxabsx = nir_fmul(b, cx, absx);
-
-   /* Y1 = B * x + C * x * fabs(x) */
-   nir_ssa_def *y1 = nir_fadd(b, bx, cxabsx);
-
-   /* Precision step: Y = P * (Y1 * fabs(Y1) - Y1) + Y1 */
-   nir_ssa_def *y = nir_fabs(b, y1);
-   y = nir_fmul(b, y, y1);
-   y = nir_fsub(b, y, y1);
-   y = nir_fmul_imm(b, y, P);
-   y = nir_fadd(b, y, y1);
-
-   return y;
-}
-
-static bool
-lower_sincos_impl(nir_function_impl *impl)
-{
-   bool progress = false;
-
-   nir_builder b;
-   nir_builder_init(&b, impl);
-
-   nir_foreach_block(block, impl) {
-      nir_foreach_instr_safe(instr, block) {
-         if (instr->type != nir_instr_type_alu)
-            continue;
-
-         nir_alu_instr *alu_instr = nir_instr_as_alu(instr);
-         nir_ssa_def *lower;
-
-         b.cursor = nir_before_instr(instr);
-
-         switch (alu_instr->op) {
-         case nir_op_fsin:
-            lower = lower_sincos(&b, nir_ssa_for_alu_src(&b, alu_instr, 0), false);
-            break;
-         case nir_op_fcos:
-            lower = lower_sincos(&b, nir_ssa_for_alu_src(&b, alu_instr, 0), true);
-            break;
-         default:
-            continue;
-         }
-
-         nir_ssa_def_rewrite_uses(&alu_instr->dest.dest.ssa,
-                                  nir_src_for_ssa(lower));
-         nir_instr_remove(instr);
-         progress = true;
-      }
-   }
-
-   if (progress) {
-      nir_metadata_preserve(impl, nir_metadata_block_index |
-                                  nir_metadata_dominance);
-   }
-
-   return progress;
-}
-
-bool
-nir_lower_sincos(nir_shader *shader)
-{
-   bool progress = false;
-
-   nir_foreach_function(function, shader) {
-      if (function->impl)
-         progress |= lower_sincos_impl(function->impl);
-   }
-
-   return progress;
-}

src/compiler/nir/nir_opt_algebraic.py

@@ -29,6 +29,7 @@ from collections import OrderedDict
 import nir_algebraic
 from nir_opcodes import type_sizes
 import itertools
+from math import pi
 
 # Convenience variables
 a = 'a'
@@ -74,6 +75,12 @@ e = 'e'
 # another condition, the two can be separated by a comma (e.g.,
 # "(many-comm-expr,is_used_once)").
 
+# based on https://web.archive.org/web/20180105155939/http://forum.devmaster.net/t/fast-and-accurate-sine-cosine/9648
+def lowered_sincos(c):
+    x = ('fsub', ('fmul', 2.0, ('ffract', ('fadd', ('fmul', 0.5 / pi, a), c))), 1.0)
+    x = ('fmul', ('fsub', x, ('fmul', x, ('fabs', x))), 4.0)
+    return ('ffma', ('ffma', x, ('fabs', x), ('fneg', x)), 0.225, x)
+
 optimizations = [
 
    (('imul', a, '#b@32(is_pos_power_of_two)'), ('ishl', a, ('find_lsb', b)), '!options->lower_bitshift'),
@@ -645,6 +652,9 @@ optimizations = [
    (('~frcp', ('fsqrt', a)), ('frsq', a)),
    (('fsqrt', a), ('frcp', ('frsq', a)), 'options->lower_fsqrt'),
    (('~frcp', ('frsq', a)), ('fsqrt', a), '!options->lower_fsqrt'),
+   # Trig
+   (('fsin', a), lowered_sincos(0.5), 'options->lower_sincos'),
+   (('fcos', a), lowered_sincos(0.75), 'options->lower_sincos'),
    # Boolean simplifications
    (('i2b32(is_used_by_if)', a), ('ine32', a, 0)),
    (('i2b1(is_used_by_if)', a), ('ine', a, 0)),

src/gallium/drivers/lima/lima_program.c

@@ -53,6 +53,7 @@ static const nir_shader_compiler_options vs_nir_options = {
    .lower_fsat = true,
    .lower_bitshift = true,
    .lower_rotate = true,
+   .lower_sincos = true,
 };
 
 static const nir_shader_compiler_options fs_nir_options = {
@@ -98,7 +99,6 @@ lima_program_optimize_vs_nir(struct nir_shader *s)
    NIR_PASS_V(s, lima_nir_lower_uniform_to_scalar);
    NIR_PASS_V(s, nir_lower_io_to_scalar,
               nir_var_shader_in|nir_var_shader_out);
-   NIR_PASS_V(s, nir_lower_sincos);
 
    do {
       progress = false;