fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-05-24 01:58:16 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

nir: update ffma helpers to use new opcodes

Browse source 
Reviewed-by: Georg Lehmann <dadschoorse@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/41165>
This commit is contained in:
Karol Herbst 2026-04-23 03:26:47 +02:00 • committed by Marge Bot
parent aeea2e7c1f
commit 109d93dd98
10 changed files with 39 additions and 39 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

6
src/amd/common/nir/ac_nir_lower_tex_coords.c
View file

@ -152,12 +152,12 @@ prepare_cube_coords(nir_builder *b, nir_tex_instr *tex, nir_def **coord, nir_src
sc = nir_fadd_imm(b, sc, 1.5);
tc = nir_fadd_imm(b, tc, 1.5);
} else {
sc = nir_ffma_imm2(b, sc, invma, 1.5);
tc = nir_ffma_imm2(b, tc, invma, 1.5);
sc = nir_ffma_weak_imm2(b, sc, invma, 1.5);
tc = nir_ffma_weak_imm2(b, tc, invma, 1.5);
}
if (tex->is_array && coords[3])
id = nir_ffma_imm1(b, coords[3], 8.0, id);
id = nir_ffma_weak_imm1(b, coords[3], 8.0, id);
*coord = nir_vec3(b, sc, tc, id);

16
src/amd/compiler/tests/test_d3d11_derivs.cpp
View file

@ -413,8 +413,8 @@ BEGIN_TEST(d3d11_derivs.cube)
pbld.add_vsfs(vs, fs);
//>> v1: %face = v_cubeid_f32 (kill)%_, (kill)%_, (kill)%_
//>> v1: %x = v_fmaak_f32 (kill)%_, %_, 0x3fc00000
//>> v1: %y = v_fmaak_f32 (kill)%_, (kill)%_, 0x3fc00000
//>> v1: %y = v_fmaak_f32 (kill)%_, %_, 0x3fc00000
//>> v1: %x = v_fmaak_f32 (kill)%_, (kill)%_, 0x3fc00000
//>> lv3: %wqm = p_start_linear_vgpr (kill)%x, (kill)%y, (kill)%face
//>> BB1
//>> v4: %_ = image_sample (kill)%_, (kill)%_, v1: undef, %wqm cube da
@ -425,8 +425,8 @@ BEGIN_TEST(d3d11_derivs.cube)
//>> v_cubeid_f32 v#rf_tmp, v#_, v#_, v#_ ; $_ $_
//>> v_mov_b32_e32 v#rf, v#rf_tmp ; $_
//>> v_fmaak_f32 v#rx_tmp, v#_, v#_, 0x3fc00000 ; $_ $_
//>> v_fmaak_f32 v#ry_tmp, v#_, v#_, 0x3fc00000 ; $_ $_
//>> v_fmaak_f32 v#rx_tmp, v#_, v#_, 0x3fc00000 ; $_ $_
//>> v_lshrrev_b64 v[#rx:#ry], 0, v[#rx_tmp:#ry_tmp] ; $_ $_
//; success = rx+1 == ry and rx+2 == rf
//>> image_sample v[#_:#_], v[#rx:#rf], s[#_:#_], s[#_:#_] dmask:0xf dim:SQ_RSRC_IMG_CUBE ; $_ $_
@ -458,10 +458,10 @@ BEGIN_TEST(d3d11_derivs.cube_array)
pbld.add_vsfs(vs, fs);
//>> v1: %face = v_cubeid_f32 (kill)%_, (kill)%_, (kill)%_
//>> v1: %x = v_fmaak_f32 (kill)%_, %_, 0x3fc00000
//>> v1: %y = v_fmaak_f32 (kill)%_, (kill)%_, 0x3fc00000
//>> v1: %layer = v_rndne_f32 (kill)%_
//>> v1: %face_layer = v_fmamk_f32 (kill)%layer, (kill)%face, 0x41000000
//>> v1: %x = v_fmaak_f32 (kill)%_, %_, 0x3fc00000
//>> v1: %y = v_fmaak_f32 (kill)%_, (kill)%_, 0x3fc00000
//>> lv3: %wqm = p_start_linear_vgpr (kill)%x, (kill)%y, (kill)%face_layer
//>> BB1
//>> v4: %_ = image_sample (kill)%_, (kill)%_, v1: undef, %wqm cube da
@ -655,8 +655,8 @@ BEGIN_TEST(d3d11_derivs.cube_txd_to_tex)
pbld.add_vsfs(vs, fs);
//>> v1: %face = v_cubeid_f32 (kill)%_, (kill)%_, (kill)%_
//>> v1: %x = v_fmaak_f32 (kill)%_, %_, 0x3fc00000
//>> v1: %y = v_fmaak_f32 (kill)%_, (kill)%_, 0x3fc00000
//>> v1: %y = v_fmaak_f32 (kill)%_, %_, 0x3fc00000
//>> v1: %x = v_fmaak_f32 (kill)%_, (kill)%_, 0x3fc00000
//>> lv3: %wqm = p_start_linear_vgpr (kill)%x, (kill)%y, (kill)%face
//>> BB1
//>> v4: %_ = image_sample (kill)%_, (kill)%_, v1: undef, %wqm cube da
@ -667,8 +667,8 @@ BEGIN_TEST(d3d11_derivs.cube_txd_to_tex)
//>> v_cubeid_f32 v#rf_tmp, v#_, v#_, v#_ ; $_ $_
//>> v_mov_b32_e32 v#rf, v#rf_tmp ; $_
//>> v_fmaak_f32 v#rx_tmp, v#_, v#_, 0x3fc00000 ; $_ $_
//>> v_fmaak_f32 v#ry_tmp, v#_, v#_, 0x3fc00000 ; $_ $_
//>> v_fmaak_f32 v#rx_tmp, v#_, v#_, 0x3fc00000 ; $_ $_
//>> v_lshrrev_b64 v[#rx:#ry], 0, v[#rx_tmp:#ry_tmp] ; $_ $_
//; success = rx+1 == ry and rx+2 == rf
//>> image_sample v[#_:#_], v[#rx:#rf], s[#_:#_], s[#_:#_] dmask:0xf dim:SQ_RSRC_IMG_CUBE ; $_ $_

16
src/compiler/nir/nir_builder.h
View file

@ -1361,33 +1361,33 @@ nir_uclamp(nir_builder *b,
}
static inline nir_def *
nir_ffma_imm12(nir_builder *build, nir_def *src0, double src1, double src2)
nir_ffma_weak_imm12(nir_builder *build, nir_def *src0, double src1, double src2)
{
if (build->shader->options &&
build->shader->options->avoid_ternary_with_two_constants)
return nir_fadd_imm(build, nir_fmul_imm(build, src0, src1), src2);
else
return nir_ffma_old(build, src0, nir_imm_floatN_t(build, src1, src0->bit_size),
nir_imm_floatN_t(build, src2, src0->bit_size));
return nir_ffma_weak(build, src0, nir_imm_floatN_t(build, src1, src0->bit_size),
nir_imm_floatN_t(build, src2, src0->bit_size));
}
static inline nir_def *
nir_ffma_imm1(nir_builder *build, nir_def *src0, double src1, nir_def *src2)
nir_ffma_weak_imm1(nir_builder *build, nir_def *src0, double src1, nir_def *src2)
{
return nir_ffma_old(build, src0, nir_imm_floatN_t(build, src1, src0->bit_size), src2);
return nir_ffma_weak(build, src0, nir_imm_floatN_t(build, src1, src0->bit_size), src2);
}
static inline nir_def *
nir_ffma_imm2(nir_builder *build, nir_def *src0, nir_def *src1, double src2)
nir_ffma_weak_imm2(nir_builder *build, nir_def *src0, nir_def *src1, double src2)
{
return nir_ffma_old(build, src0, src1, nir_imm_floatN_t(build, src2, src0->bit_size));
return nir_ffma_weak(build, src0, src1, nir_imm_floatN_t(build, src2, src0->bit_size));
}
static inline nir_def *
nir_a_minus_bc(nir_builder *build, nir_def *src0, nir_def *src1,
nir_def *src2)
{
return nir_ffma_old(build, nir_fneg(build, src1), src2, src0);
return nir_ffma_weak(build, nir_fneg(build, src1), src2, src0);
}
static inline nir_def *

14
src/compiler/nir/nir_builtin_builder.c
View file

@ -199,12 +199,12 @@ build_asin(nir_builder *b, nir_def *x, bool piecewise)
}
nir_def *abs_x = nir_fabs(b, x);
nir_def *p0_plus_xp1 = nir_ffma_imm12(b, abs_x, -0.0187293, 0.0742610);
nir_def *p0_plus_xp1 = nir_ffma_weak_imm12(b, abs_x, -0.0187293, 0.0742610);
nir_def *expr_tail =
nir_ffma_imm2(b, abs_x,
nir_ffma_imm2(b, abs_x, p0_plus_xp1, -0.2121144),
1.5707288);
nir_ffma_weak_imm2(b, abs_x,
nir_ffma_weak_imm2(b, abs_x, p0_plus_xp1, -0.2121144),
1.5707288);
nir_def *result0 = nir_fmul(b, nir_fsign(b, x),
nir_a_minus_bc(b, nir_imm_floatN_t(b, M_PI_2f, x->bit_size),
@ -217,7 +217,7 @@ build_asin(nir_builder *b, nir_def *x, bool piecewise)
nir_def *x2 = nir_fmul(b, x, x);
nir_def *result1 = nir_fmul(b,
x,
nir_ffma_imm12(b, x2, (1.0/6.0), 1.0));
nir_ffma_weak_imm12(b, x2, (1.0/6.0), 1.0));
return nir_bcsel(b,
nir_flt_imm(b, abs_x, 0.21502245),
result1,
@ -277,7 +277,7 @@ nir_atan(nir_builder *b, nir_def *y_over_x)
nir_def *res = nir_imm_floatN_t(b, coeffs[0], bit_size);
for (unsigned i = 1; i < ARRAY_SIZE(coeffs); ++i) {
res = nir_ffma_imm2(b, res, x_2, coeffs[i]);
res = nir_ffma_weak_imm2(b, res, x_2, coeffs[i]);
}
/* range-reduction fixup value */
@ -359,7 +359,7 @@ nir_atan2(nir_builder *b, nir_def *y, nir_def *x)
* coordinate system.
*/
nir_def *arc =
nir_ffma_imm1(b, nir_b2fN(b, flip, bit_size), M_PI_2, nir_atan(b, tan));
nir_ffma_weak_imm1(b, nir_b2fN(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

2
src/compiler/nir/nir_opt_varyings.c
View file

@ -1075,7 +1075,7 @@ static nir_def *
build_convert_inf_to_nan(nir_builder *b, nir_def *x)
{
/* Do x*0 + x. The multiplication by 0 can't be optimized out. */
nir_def *fma = nir_ffma_imm1(b, x, 0, x);
nir_def *fma = nir_ffma_weak_imm1(b, x, 0, x);
nir_def_as_alu(fma)->fp_math_ctrl = nir_fp_preserve_nan | nir_fp_preserve_inf | nir_fp_exact;
return fma;
}

2
src/compiler/spirv/vtn_amd.c
View file

@ -20,7 +20,7 @@ vtn_handle_amd_gcn_shader_instruction(struct vtn_builder *b, SpvOp ext_opcode,
def = nir_cube_amd(&b->nb, vtn_get_nir_ssa(b, w[5]));
nir_def *st = nir_swizzle(&b->nb, def, (unsigned[]){1, 0}, 2);
nir_def *invma = nir_frcp(&b->nb, nir_channel(&b->nb, def, 2));
def = nir_ffma_imm2(&b->nb, st, invma, 0.5);
def = nir_ffma_weak_imm2(&b->nb, st, invma, 0.5);
break;
}
case TimeAMD: {

4
src/compiler/spirv/vtn_glsl450.c
View file

@ -503,11 +503,11 @@ handle_glsl450_alu(struct vtn_builder *b, enum GLSLstd450 entrypoint,
case GLSLstd450Asinh:
dest->def = nir_fmul(nb, nir_fsign(nb, src[0]),
nir_flog(nb, nir_fadd(nb, nir_fabs(nb, src[0]),
nir_fsqrt(nb, nir_ffma_imm2(nb, src[0], src[0], 1.0f)))));
nir_fsqrt(nb, nir_ffma_weak_imm2(nb, src[0], src[0], 1.0f)))));
break;
case GLSLstd450Acosh:
dest->def = nir_flog(nb, nir_fadd(nb, src[0],
nir_fsqrt(nb, nir_ffma_imm2(nb, src[0], src[0], -1.0f))));
nir_fsqrt(nb, nir_ffma_weak_imm2(nb, src[0], src[0], -1.0f))));
break;
case GLSLstd450Atanh: {
dest->def =

2
src/gallium/auxiliary/vl/vl_compositor_cs.c
View file

@ -214,7 +214,7 @@ static inline nir_def *cs_chroma_offset(struct cs_shader *s, nir_def *src, unsig
nir_def *offset = nir_channels(b, s->params[3], 0x3 << 2);
if (flags & COORDS_CHROMA)
return nir_fadd(b, src, offset);
return nir_ffma_imm1(b, offset, -0.5f, src);
return nir_ffma_weak_imm1(b, offset, -0.5f, src);
}
static inline nir_def *cs_clamp(struct cs_shader *s, nir_def *src, unsigned flags)

10
src/gallium/drivers/r600/sfn/sfn_nir_lower_alu.cpp
View file

@ -91,15 +91,15 @@ LowerSinCos::lower(nir_instr *instr)
assert(alu->op == nir_op_fsin || alu->op == nir_op_fcos);
auto fract = nir_ffract(b,
nir_ffma_imm12(b,
nir_ssa_for_alu_src(b, alu, 0),
0.15915494,
0.5));
nir_ffma_weak_imm12(b,
nir_ssa_for_alu_src(b, alu, 0),
0.15915494,
0.5));
auto normalized =
m_gxf_level != R600
? nir_fadd_imm(b, fract, -0.5)
: nir_ffma_imm12(b, fract, 2.0f * M_PI, -M_PI);
: nir_ffma_weak_imm12(b, fract, 2.0f * M_PI, -M_PI);
if (alu->op == nir_op_fsin)
return nir_fsin_normalized_2_pi(b, normalized);

6
src/vulkan/runtime/vk_nir_convert_ycbcr.c
View file

@ -81,9 +81,9 @@ nir_convert_ycbcr_to_rgb(nir_builder *b,
nir_def *expanded_channels =
nir_vec4(b,
nir_ffma_imm12(b, nir_channel(b, raw_channels, 0), range_coeffs[2][0], range_coeffs[2][1]),
nir_ffma_imm12(b, nir_channel(b, raw_channels, 1), range_coeffs[0][0], range_coeffs[0][1]),
nir_ffma_imm12(b, nir_channel(b, raw_channels, 2), range_coeffs[1][0], range_coeffs[1][1]),
nir_ffma_weak_imm12(b, nir_channel(b, raw_channels, 0), range_coeffs[2][0], range_coeffs[2][1]),
nir_ffma_weak_imm12(b, nir_channel(b, raw_channels, 1), range_coeffs[0][0], range_coeffs[0][1]),
nir_ffma_weak_imm12(b, nir_channel(b, raw_channels, 2), range_coeffs[1][0], range_coeffs[1][1]),
nir_channel(b, raw_channels, 3));
if (model == VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_IDENTITY)