From: Neil Roberts Date: Fri, 11 Oct 2019 13:43:47 +0000 (+0200) Subject: nir/builder: Move nir_atan and nir_atan2 from SPIR-V translator X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=2098ae16c8b4e64d0694a28f74a461b69b98a336;p=mesa.git nir/builder: Move nir_atan and nir_atan2 from SPIR-V translator Moves build_atan and build_atan2 into nir_builtin_builder. The goal is to be able to use this from the GLSL translator too. Reviewed-by: Kristian H. Kristensen --- diff --git a/src/compiler/nir/nir_builtin_builder.c b/src/compiler/nir/nir_builtin_builder.c index 8812c7a6a8a..c27666949e9 100644 --- a/src/compiler/nir/nir_builtin_builder.c +++ b/src/compiler/nir/nir_builtin_builder.c @@ -1,5 +1,6 @@ /* * Copyright © 2018 Red Hat Inc. + * Copyright © 2015 Intel Corporation * * Permission is hereby granted, free of charge, to any person obtaining a * copy of this software and associated documentation files (the "Software"), @@ -173,3 +174,154 @@ nir_upsample(nir_builder *b, nir_ssa_def *hi, nir_ssa_def *lo) return nir_vec(b, res, lo->num_components); } + +/** + * Compute xs[0] + xs[1] + xs[2] + ... using fadd. + */ +static nir_ssa_def * +build_fsum(nir_builder *b, nir_ssa_def **xs, int terms) +{ + nir_ssa_def *accum = xs[0]; + + for (int i = 1; i < terms; i++) + accum = nir_fadd(b, accum, xs[i]); + + return accum; +} + +nir_ssa_def * +nir_atan(nir_builder *b, nir_ssa_def *y_over_x) +{ + const uint32_t bit_size = y_over_x->bit_size; + + nir_ssa_def *abs_y_over_x = nir_fabs(b, y_over_x); + nir_ssa_def *one = nir_imm_floatN_t(b, 1.0f, bit_size); + + /* + * range-reduction, first step: + * + * / y_over_x if |y_over_x| <= 1.0; + * x = < + * \ 1.0 / y_over_x otherwise + */ + nir_ssa_def *x = nir_fdiv(b, nir_fmin(b, abs_y_over_x, one), + nir_fmax(b, abs_y_over_x, one)); + + /* + * approximate atan by evaluating polynomial: + * + * x * 0.9999793128310355 - x^3 * 0.3326756418091246 + + * x^5 * 0.1938924977115610 - x^7 * 0.1173503194786851 + + * x^9 * 0.0536813784310406 - x^11 * 0.0121323213173444 + */ + nir_ssa_def *x_2 = nir_fmul(b, x, x); + nir_ssa_def *x_3 = nir_fmul(b, x_2, x); + nir_ssa_def *x_5 = nir_fmul(b, x_3, x_2); + nir_ssa_def *x_7 = nir_fmul(b, x_5, x_2); + nir_ssa_def *x_9 = nir_fmul(b, x_7, x_2); + nir_ssa_def *x_11 = nir_fmul(b, x_9, x_2); + + nir_ssa_def *polynomial_terms[] = { + nir_fmul_imm(b, x, 0.9999793128310355f), + nir_fmul_imm(b, x_3, -0.3326756418091246f), + nir_fmul_imm(b, x_5, 0.1938924977115610f), + nir_fmul_imm(b, x_7, -0.1173503194786851f), + nir_fmul_imm(b, x_9, 0.0536813784310406f), + nir_fmul_imm(b, x_11, -0.0121323213173444f), + }; + + nir_ssa_def *tmp = + build_fsum(b, polynomial_terms, ARRAY_SIZE(polynomial_terms)); + + /* range-reduction fixup */ + tmp = nir_fadd(b, tmp, + nir_fmul(b, nir_b2f(b, nir_flt(b, one, abs_y_over_x), bit_size), + nir_fadd_imm(b, nir_fmul_imm(b, tmp, -2.0f), M_PI_2))); + + /* sign fixup */ + return nir_fmul(b, tmp, nir_fsign(b, y_over_x)); +} + +nir_ssa_def * +nir_atan2(nir_builder *b, nir_ssa_def *y, nir_ssa_def *x) +{ + assert(y->bit_size == x->bit_size); + const uint32_t bit_size = x->bit_size; + + nir_ssa_def *zero = nir_imm_floatN_t(b, 0, bit_size); + nir_ssa_def *one = nir_imm_floatN_t(b, 1, bit_size); + + /* If we're on the left half-plane rotate the coordinates π/2 clock-wise + * for the y=0 discontinuity to end up aligned with the vertical + * discontinuity of atan(s/t) along t=0. This also makes sure that we + * don't attempt to divide by zero along the vertical line, which may give + * unspecified results on non-GLSL 4.1-capable hardware. + */ + nir_ssa_def *flip = nir_fge(b, zero, x); + nir_ssa_def *s = nir_bcsel(b, flip, nir_fabs(b, x), y); + nir_ssa_def *t = nir_bcsel(b, flip, y, nir_fabs(b, x)); + + /* If the magnitude of the denominator exceeds some huge value, scale down + * the arguments in order to prevent the reciprocal operation from flushing + * its result to zero, which would cause precision problems, and for s + * infinite would cause us to return a NaN instead of the correct finite + * value. + * + * If fmin and fmax are respectively the smallest and largest positive + * normalized floating point values representable by the implementation, + * the constants below should be in agreement with: + * + * huge <= 1 / fmin + * scale <= 1 / fmin / fmax (for |t| >= huge) + * + * In addition scale should be a negative power of two in order to avoid + * loss of precision. The values chosen below should work for most usual + * floating point representations with at least the dynamic range of ATI's + * 24-bit representation. + */ + const double huge_val = bit_size >= 32 ? 1e18 : 16384; + nir_ssa_def *huge = nir_imm_floatN_t(b, huge_val, bit_size); + nir_ssa_def *scale = nir_bcsel(b, nir_fge(b, nir_fabs(b, t), huge), + nir_imm_floatN_t(b, 0.25, bit_size), one); + nir_ssa_def *rcp_scaled_t = nir_frcp(b, nir_fmul(b, t, scale)); + nir_ssa_def *s_over_t = nir_fmul(b, nir_fmul(b, s, scale), rcp_scaled_t); + + /* For |x| = |y| assume tan = 1 even if infinite (i.e. pretend momentarily + * that ∞/∞ = 1) in order to comply with the rather artificial rules + * inherited from IEEE 754-2008, namely: + * + * "atan2(±∞, −∞) is ±3π/4 + * atan2(±∞, +∞) is ±π/4" + * + * Note that this is inconsistent with the rules for the neighborhood of + * zero that are based on iterated limits: + * + * "atan2(±0, −0) is ±π + * atan2(±0, +0) is ±0" + * + * but GLSL specifically allows implementations to deviate from IEEE rules + * at (0,0), so we take that license (i.e. pretend that 0/0 = 1 here as + * well). + */ + nir_ssa_def *tan = nir_bcsel(b, nir_feq(b, nir_fabs(b, x), nir_fabs(b, y)), + one, nir_fabs(b, s_over_t)); + + /* Calculate the arctangent and fix up the result if we had flipped the + * coordinate system. + */ + nir_ssa_def *arc = + nir_fadd(b, nir_fmul_imm(b, nir_b2f(b, flip, bit_size), M_PI_2), + nir_atan(b, tan)); + + /* Rather convoluted calculation of the sign of the result. When x < 0 we + * cannot use fsign because we need to be able to distinguish between + * negative and positive zero. We don't use bitwise arithmetic tricks for + * consistency with the GLSL front-end. When x >= 0 rcp_scaled_t will + * always be non-negative so this won't be able to distinguish between + * negative and positive zero, but we don't care because atan2 is + * continuous along the whole positive y = 0 half-line, so it won't affect + * the result significantly. + */ + return nir_bcsel(b, nir_flt(b, nir_fmin(b, y, rcp_scaled_t), zero), + nir_fneg(b, arc), arc); +} diff --git a/src/compiler/nir/nir_builtin_builder.h b/src/compiler/nir/nir_builtin_builder.h index 39a1e0bd3cd..519938f8bfe 100644 --- a/src/compiler/nir/nir_builtin_builder.h +++ b/src/compiler/nir/nir_builtin_builder.h @@ -41,6 +41,8 @@ nir_ssa_def* nir_rotate(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y); nir_ssa_def* nir_smoothstep(nir_builder *b, nir_ssa_def *edge0, nir_ssa_def *edge1, nir_ssa_def *x); nir_ssa_def* nir_upsample(nir_builder *b, nir_ssa_def *hi, nir_ssa_def *lo); +nir_ssa_def* nir_atan(nir_builder *b, nir_ssa_def *y_over_x); +nir_ssa_def* nir_atan2(nir_builder *b, nir_ssa_def *y, nir_ssa_def *x); static inline nir_ssa_def * nir_nan_check2(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y, nir_ssa_def *res) diff --git a/src/compiler/spirv/vtn_glsl450.c b/src/compiler/spirv/vtn_glsl450.c index dd72a86e21c..2d66512bc42 100644 --- a/src/compiler/spirv/vtn_glsl450.c +++ b/src/compiler/spirv/vtn_glsl450.c @@ -234,157 +234,6 @@ build_asin(nir_builder *b, nir_ssa_def *x, float p0, float p1) expr_tail))); } -/** - * Compute xs[0] + xs[1] + xs[2] + ... using fadd. - */ -static nir_ssa_def * -build_fsum(nir_builder *b, nir_ssa_def **xs, int terms) -{ - nir_ssa_def *accum = xs[0]; - - for (int i = 1; i < terms; i++) - accum = nir_fadd(b, accum, xs[i]); - - return accum; -} - -static nir_ssa_def * -build_atan(nir_builder *b, nir_ssa_def *y_over_x) -{ - const uint32_t bit_size = y_over_x->bit_size; - - nir_ssa_def *abs_y_over_x = nir_fabs(b, y_over_x); - nir_ssa_def *one = nir_imm_floatN_t(b, 1.0f, bit_size); - - /* - * range-reduction, first step: - * - * / y_over_x if |y_over_x| <= 1.0; - * x = < - * \ 1.0 / y_over_x otherwise - */ - nir_ssa_def *x = nir_fdiv(b, nir_fmin(b, abs_y_over_x, one), - nir_fmax(b, abs_y_over_x, one)); - - /* - * approximate atan by evaluating polynomial: - * - * x * 0.9999793128310355 - x^3 * 0.3326756418091246 + - * x^5 * 0.1938924977115610 - x^7 * 0.1173503194786851 + - * x^9 * 0.0536813784310406 - x^11 * 0.0121323213173444 - */ - nir_ssa_def *x_2 = nir_fmul(b, x, x); - nir_ssa_def *x_3 = nir_fmul(b, x_2, x); - nir_ssa_def *x_5 = nir_fmul(b, x_3, x_2); - nir_ssa_def *x_7 = nir_fmul(b, x_5, x_2); - nir_ssa_def *x_9 = nir_fmul(b, x_7, x_2); - nir_ssa_def *x_11 = nir_fmul(b, x_9, x_2); - - nir_ssa_def *polynomial_terms[] = { - nir_fmul_imm(b, x, 0.9999793128310355f), - nir_fmul_imm(b, x_3, -0.3326756418091246f), - nir_fmul_imm(b, x_5, 0.1938924977115610f), - nir_fmul_imm(b, x_7, -0.1173503194786851f), - nir_fmul_imm(b, x_9, 0.0536813784310406f), - nir_fmul_imm(b, x_11, -0.0121323213173444f), - }; - - nir_ssa_def *tmp = - build_fsum(b, polynomial_terms, ARRAY_SIZE(polynomial_terms)); - - /* range-reduction fixup */ - tmp = nir_fadd(b, tmp, - nir_fmul(b, nir_b2f(b, nir_flt(b, one, abs_y_over_x), bit_size), - nir_fadd_imm(b, nir_fmul_imm(b, tmp, -2.0f), M_PI_2f))); - - /* sign fixup */ - return nir_fmul(b, tmp, nir_fsign(b, y_over_x)); -} - -static nir_ssa_def * -build_atan2(nir_builder *b, nir_ssa_def *y, nir_ssa_def *x) -{ - assert(y->bit_size == x->bit_size); - const uint32_t bit_size = x->bit_size; - - nir_ssa_def *zero = nir_imm_floatN_t(b, 0, bit_size); - nir_ssa_def *one = nir_imm_floatN_t(b, 1, bit_size); - - /* If we're on the left half-plane rotate the coordinates π/2 clock-wise - * for the y=0 discontinuity to end up aligned with the vertical - * discontinuity of atan(s/t) along t=0. This also makes sure that we - * don't attempt to divide by zero along the vertical line, which may give - * unspecified results on non-GLSL 4.1-capable hardware. - */ - nir_ssa_def *flip = nir_fge(b, zero, x); - nir_ssa_def *s = nir_bcsel(b, flip, nir_fabs(b, x), y); - nir_ssa_def *t = nir_bcsel(b, flip, y, nir_fabs(b, x)); - - /* If the magnitude of the denominator exceeds some huge value, scale down - * the arguments in order to prevent the reciprocal operation from flushing - * its result to zero, which would cause precision problems, and for s - * infinite would cause us to return a NaN instead of the correct finite - * value. - * - * If fmin and fmax are respectively the smallest and largest positive - * normalized floating point values representable by the implementation, - * the constants below should be in agreement with: - * - * huge <= 1 / fmin - * scale <= 1 / fmin / fmax (for |t| >= huge) - * - * In addition scale should be a negative power of two in order to avoid - * loss of precision. The values chosen below should work for most usual - * floating point representations with at least the dynamic range of ATI's - * 24-bit representation. - */ - const double huge_val = bit_size >= 32 ? 1e18 : 16384; - nir_ssa_def *huge = nir_imm_floatN_t(b, huge_val, bit_size); - nir_ssa_def *scale = nir_bcsel(b, nir_fge(b, nir_fabs(b, t), huge), - nir_imm_floatN_t(b, 0.25, bit_size), one); - nir_ssa_def *rcp_scaled_t = nir_frcp(b, nir_fmul(b, t, scale)); - nir_ssa_def *s_over_t = nir_fmul(b, nir_fmul(b, s, scale), rcp_scaled_t); - - /* For |x| = |y| assume tan = 1 even if infinite (i.e. pretend momentarily - * that ∞/∞ = 1) in order to comply with the rather artificial rules - * inherited from IEEE 754-2008, namely: - * - * "atan2(±∞, −∞) is ±3π/4 - * atan2(±∞, +∞) is ±π/4" - * - * Note that this is inconsistent with the rules for the neighborhood of - * zero that are based on iterated limits: - * - * "atan2(±0, −0) is ±π - * atan2(±0, +0) is ±0" - * - * but GLSL specifically allows implementations to deviate from IEEE rules - * at (0,0), so we take that license (i.e. pretend that 0/0 = 1 here as - * well). - */ - nir_ssa_def *tan = nir_bcsel(b, nir_feq(b, nir_fabs(b, x), nir_fabs(b, y)), - one, nir_fabs(b, s_over_t)); - - /* Calculate the arctangent and fix up the result if we had flipped the - * coordinate system. - */ - nir_ssa_def *arc = - nir_fadd(b, nir_fmul_imm(b, nir_b2f(b, flip, bit_size), M_PI_2f), - build_atan(b, tan)); - - /* Rather convoluted calculation of the sign of the result. When x < 0 we - * cannot use fsign because we need to be able to distinguish between - * negative and positive zero. We don't use bitwise arithmetic tricks for - * consistency with the GLSL front-end. When x >= 0 rcp_scaled_t will - * always be non-negative so this won't be able to distinguish between - * negative and positive zero, but we don't care because atan2 is - * continuous along the whole positive y = 0 half-line, so it won't affect - * the result significantly. - */ - return nir_bcsel(b, nir_flt(b, nir_fmin(b, y, rcp_scaled_t), zero), - nir_fneg(b, arc), arc); -} - static nir_op vtn_nir_alu_op_for_spirv_glsl_opcode(struct vtn_builder *b, enum GLSLstd450 opcode, @@ -662,11 +511,11 @@ handle_glsl450_alu(struct vtn_builder *b, enum GLSLstd450 entrypoint, return; case GLSLstd450Atan: - val->ssa->def = build_atan(nb, src[0]); + val->ssa->def = nir_atan(nb, src[0]); return; case GLSLstd450Atan2: - val->ssa->def = build_atan2(nb, src[0], src[1]); + val->ssa->def = nir_atan2(nb, src[0], src[1]); return; case GLSLstd450Frexp: {