From b38d3cdceab07803a7dd81f8870a892fd82a077c Mon Sep 17 00:00:00 2001 From: Arcady Goldmints-Orlov Date: Wed, 15 Apr 2020 16:55:53 -0500 Subject: [PATCH] nir/spirv/glsl450: increase asin(x) precision asin(x) is now implemented using a piecewise approximation, which improves the precision for |x| < 0.5 Previously, we were using a polynomial approximation for both the asin() and acos() functions. Unfortunately, for asin(), this polynomial does not have enough precision to satisfy the Vulkan CTS requiremenents, which define the asin() precision based on the precision of atan2(x, sqrt(1.0 - x*x)). The piecewise approximation gives the needed precision in the problematic range. v2: Skip the piecewise approximation for acos Closes: #1843 Acked-by: Francisco Jerez Part-of: --- .gitlab-ci/deqp-radv-fiji-aco-fails.txt | 6 --- .gitlab-ci/deqp-radv-navi10-aco-fails.txt | 6 --- .gitlab-ci/deqp-radv-pitcairn-aco-fails.txt | 6 --- .gitlab-ci/deqp-radv-polaris10-aco-fails.txt | 6 --- .gitlab-ci/deqp-radv-raven-aco-fails.txt | 6 --- .gitlab-ci/deqp-radv-vega10-aco-fails.txt | 6 --- src/compiler/spirv/vtn_glsl450.c | 42 +++++++++++++++----- 7 files changed, 33 insertions(+), 45 deletions(-) diff --git a/.gitlab-ci/deqp-radv-fiji-aco-fails.txt b/.gitlab-ci/deqp-radv-fiji-aco-fails.txt index 3754ad5e9fc..a75e20aca18 100644 --- a/.gitlab-ci/deqp-radv-fiji-aco-fails.txt +++ b/.gitlab-ci/deqp-radv-fiji-aco-fails.txt @@ -2,12 +2,6 @@ dEQP-VK.api.buffer_marker.graphics.default_mem.bottom_of_pipe.memory_dep.draw dEQP-VK.api.buffer_marker.graphics.external_host_mem.bottom_of_pipe.memory_dep.draw -# SPIR-V->NIR issues. -dEQP-VK.glsl.builtin.precision.asin.highp.scalar -dEQP-VK.glsl.builtin.precision.asin.highp.vec2 -dEQP-VK.glsl.builtin.precision.asin.highp.vec3 -dEQP-VK.glsl.builtin.precision.asin.highp.vec4 - # CTS bug (list of extensions not up-to-date). dEQP-VK.info.device_extensions diff --git a/.gitlab-ci/deqp-radv-navi10-aco-fails.txt b/.gitlab-ci/deqp-radv-navi10-aco-fails.txt index 39b0e5bbabb..c3f80306bf7 100644 --- a/.gitlab-ci/deqp-radv-navi10-aco-fails.txt +++ b/.gitlab-ci/deqp-radv-navi10-aco-fails.txt @@ -1,9 +1,3 @@ -# SPIR-V->NIR issues. -dEQP-VK.glsl.builtin.precision.asin.highp.scalar -dEQP-VK.glsl.builtin.precision.asin.highp.vec2 -dEQP-VK.glsl.builtin.precision.asin.highp.vec3 -dEQP-VK.glsl.builtin.precision.asin.highp.vec4 - # CTS bug (list of extensions not up-to-date). dEQP-VK.info.device_extensions diff --git a/.gitlab-ci/deqp-radv-pitcairn-aco-fails.txt b/.gitlab-ci/deqp-radv-pitcairn-aco-fails.txt index 7541df0c183..96b78d9f613 100644 --- a/.gitlab-ci/deqp-radv-pitcairn-aco-fails.txt +++ b/.gitlab-ci/deqp-radv-pitcairn-aco-fails.txt @@ -1,8 +1,2 @@ -# SPIR-V->NIR issues. -dEQP-VK.glsl.builtin.precision.asin.highp.scalar -dEQP-VK.glsl.builtin.precision.asin.highp.vec2 -dEQP-VK.glsl.builtin.precision.asin.highp.vec3 -dEQP-VK.glsl.builtin.precision.asin.highp.vec4 - # CTS bug (list of extensions not up-to-date). dEQP-VK.info.device_extensions diff --git a/.gitlab-ci/deqp-radv-polaris10-aco-fails.txt b/.gitlab-ci/deqp-radv-polaris10-aco-fails.txt index ae8d825f416..8752be3b83b 100644 --- a/.gitlab-ci/deqp-radv-polaris10-aco-fails.txt +++ b/.gitlab-ci/deqp-radv-polaris10-aco-fails.txt @@ -2,11 +2,5 @@ dEQP-VK.api.buffer_marker.graphics.default_mem.bottom_of_pipe.memory_dep.draw dEQP-VK.api.buffer_marker.graphics.external_host_mem.bottom_of_pipe.memory_dep.draw -# SPIR-V->NIR issues. -dEQP-VK.glsl.builtin.precision.asin.highp.scalar -dEQP-VK.glsl.builtin.precision.asin.highp.vec2 -dEQP-VK.glsl.builtin.precision.asin.highp.vec3 -dEQP-VK.glsl.builtin.precision.asin.highp.vec4 - # CTS bug (list of extensions not up-to-date). dEQP-VK.info.device_extensions diff --git a/.gitlab-ci/deqp-radv-raven-aco-fails.txt b/.gitlab-ci/deqp-radv-raven-aco-fails.txt index 39b0e5bbabb..c3f80306bf7 100644 --- a/.gitlab-ci/deqp-radv-raven-aco-fails.txt +++ b/.gitlab-ci/deqp-radv-raven-aco-fails.txt @@ -1,9 +1,3 @@ -# SPIR-V->NIR issues. -dEQP-VK.glsl.builtin.precision.asin.highp.scalar -dEQP-VK.glsl.builtin.precision.asin.highp.vec2 -dEQP-VK.glsl.builtin.precision.asin.highp.vec3 -dEQP-VK.glsl.builtin.precision.asin.highp.vec4 - # CTS bug (list of extensions not up-to-date). dEQP-VK.info.device_extensions diff --git a/.gitlab-ci/deqp-radv-vega10-aco-fails.txt b/.gitlab-ci/deqp-radv-vega10-aco-fails.txt index 39b0e5bbabb..c3f80306bf7 100644 --- a/.gitlab-ci/deqp-radv-vega10-aco-fails.txt +++ b/.gitlab-ci/deqp-radv-vega10-aco-fails.txt @@ -1,9 +1,3 @@ -# SPIR-V->NIR issues. -dEQP-VK.glsl.builtin.precision.asin.highp.scalar -dEQP-VK.glsl.builtin.precision.asin.highp.vec2 -dEQP-VK.glsl.builtin.precision.asin.highp.vec3 -dEQP-VK.glsl.builtin.precision.asin.highp.vec4 - # CTS bug (list of extensions not up-to-date). dEQP-VK.info.device_extensions diff --git a/src/compiler/spirv/vtn_glsl450.c b/src/compiler/spirv/vtn_glsl450.c index ca836bde61b..947d33c6bf6 100644 --- a/src/compiler/spirv/vtn_glsl450.c +++ b/src/compiler/spirv/vtn_glsl450.c @@ -172,17 +172,18 @@ matrix_inverse(struct vtn_builder *b, struct vtn_ssa_value *src) } /** - * Approximate asin(x) by the formula: - * asin~(x) = sign(x) * (pi/2 - sqrt(1 - |x|) * (pi/2 + |x|(pi/4 - 1 + |x|(p0 + |x|p1)))) + * Approximate asin(x) by the piecewise formula: + * for |x| < 0.5, asin~(x) = x * (1 + x²(pS0 + x²(pS1 + x²*pS2)) / (1 + x²*qS1)) + * for |x| ≥ 0.5, asin~(x) = sign(x) * (π/2 - sqrt(1 - |x|) * (π/2 + |x|(π/4 - 1 + |x|(p0 + |x|p1)))) * - * which is correct to first order at x=0 and x=±1 regardless of the p + * The latter is correct to first order at x=0 and x=±1 regardless of the p * coefficients but can be made second-order correct at both ends by selecting * the fit coefficients appropriately. Different p coefficients can be used * in the asin and acos implementation to minimize some relative error metric * in each case. */ static nir_ssa_def * -build_asin(nir_builder *b, nir_ssa_def *x, float p0, float p1) +build_asin(nir_builder *b, nir_ssa_def *x, float p0, float p1, bool piecewise) { if (x->bit_size == 16) { /* The polynomial approximation isn't precise enough to meet half-float @@ -195,10 +196,10 @@ build_asin(nir_builder *b, nir_ssa_def *x, float p0, float p1) * approximation in 32-bit math and then we convert the result back to * 16-bit. */ - return nir_f2f16(b, build_asin(b, nir_f2f32(b, x), p0, p1)); + return nir_f2f16(b, build_asin(b, nir_f2f32(b, x), p0, p1, piecewise)); } - nir_ssa_def *one = nir_imm_floatN_t(b, 1.0f, x->bit_size); + nir_ssa_def *half = nir_imm_floatN_t(b, 0.5f, x->bit_size); nir_ssa_def *abs_x = nir_fabs(b, x); nir_ssa_def *p0_plus_xp1 = nir_fadd_imm(b, nir_fmul_imm(b, abs_x, p1), p0); @@ -210,10 +211,33 @@ build_asin(nir_builder *b, nir_ssa_def *x, float p0, float p1) M_PI_4f - 1.0f)), M_PI_2f); - return nir_fmul(b, nir_fsign(b, x), + nir_ssa_def *result0 = nir_fmul(b, nir_fsign(b, x), nir_fsub(b, nir_imm_floatN_t(b, M_PI_2f, x->bit_size), nir_fmul(b, nir_fsqrt(b, nir_fsub(b, one, abs_x)), expr_tail))); + if (piecewise) { + /* approximation for |x| < 0.5 */ + const float pS0 = 1.6666586697e-01f; + const float pS1 = -4.2743422091e-02f; + const float pS2 = -8.6563630030e-03f; + const float qS1 = -7.0662963390e-01f; + + nir_ssa_def *x2 = nir_fmul(b, x, x); + nir_ssa_def *p = nir_fmul(b, + x2, + nir_fadd_imm(b, + nir_fmul(b, + x2, + nir_fadd_imm(b, nir_fmul_imm(b, x2, pS2), + pS1)), + pS0)); + + nir_ssa_def *q = nir_fadd(b, one, nir_fmul_imm(b, x2, qS1)); + nir_ssa_def *result1 = nir_fadd(b, x, nir_fmul(b, x, nir_fdiv(b, p, q))); + return nir_bcsel(b, nir_flt(b, abs_x, half), result1, result0); + } else { + return result0; + } } static nir_op @@ -487,13 +511,13 @@ handle_glsl450_alu(struct vtn_builder *b, enum GLSLstd450 entrypoint, } case GLSLstd450Asin: - val->ssa->def = build_asin(nb, src[0], 0.086566724, -0.03102955); + val->ssa->def = build_asin(nb, src[0], 0.086566724, -0.03102955, true); return; case GLSLstd450Acos: val->ssa->def = nir_fsub(nb, nir_imm_floatN_t(nb, M_PI_2f, src[0]->bit_size), - build_asin(nb, src[0], 0.08132463, -0.02363318)); + build_asin(nb, src[0], 0.08132463, -0.02363318, false)); return; case GLSLstd450Atan: -- 2.30.2