/third_party/optimized-routines/math/ |
H A D | v_expf.c | 41 v_u32_t r1 = v_as_u32_f32 (v_fma_f32 (poly, s2, s2) * s1); in specialcase() 43 v_u32_t r0 = v_as_u32_f32 (v_fma_f32 (poly, scale, scale)); in specialcase() 57 z = v_fma_f32 (x, InvLn2, Shift); in expf() 59 r = v_fma_f32 (n, -Ln2hi, x); in expf() 60 r = v_fma_f32 (n, -Ln2lo, r); in expf() 65 r = v_fma_f32 (n, -Ln2hi, x); in expf() 66 r = v_fma_f32 (n, -Ln2lo, r); in expf() 73 p = v_fma_f32 (C0, r, C1); in expf() 74 q = v_fma_f32 (C2, r, C3); in expf() 75 q = v_fma_f32 ( in expf() [all...] |
H A D | v_expf_1u.c | 55 z = v_fma_f32 (x, InvLn2, Shift); in expf_1u() 57 r = v_fma_f32 (n, -Ln2hi, x); in expf_1u() 58 r = v_fma_f32 (n, -Ln2lo, r); in expf_1u() 63 r = v_fma_f32 (n, -Ln2hi, x); in expf_1u() 64 r = v_fma_f32 (n, -Ln2lo, r); in expf_1u() 70 poly = v_fma_f32 (C0, r, C1); in expf_1u() 71 poly = v_fma_f32 (poly, r, C2); in expf_1u() 72 poly = v_fma_f32 (poly, r, C3); in expf_1u() 73 poly = v_fma_f32 (poly, r, C4); in expf_1u() 74 poly = v_fma_f32 (pol in expf_1u() [all...] |
H A D | v_cosf.c | 51 n = v_fma_f32 (InvPi, r + HalfPi, Shift); in cosf() 57 r = v_fma_f32 (-Pi1, n, r); in cosf() 58 r = v_fma_f32 (-Pi2, n, r); in cosf() 59 r = v_fma_f32 (-Pi3, n, r); in cosf() 63 y = v_fma_f32 (A9, r2, A7); in cosf() 64 y = v_fma_f32 (y, r2, A5); in cosf() 65 y = v_fma_f32 (y, r2, A3); in cosf() 66 y = v_fma_f32 (y * r2, r, r); in cosf()
|
H A D | v_logf.c | 59 p = v_fma_f32 (P6, r, P5); in logf() 60 q = v_fma_f32 (P4, r, P3); in logf() 61 y = v_fma_f32 (P2, r, P1); in logf() 62 p = v_fma_f32 (P7, r2, p); in logf() 63 q = v_fma_f32 (p, r2, q); in logf() 64 y = v_fma_f32 (q, r2, y); in logf() 65 p = v_fma_f32 (Ln2, n, r); in logf() 66 y = v_fma_f32 (y, r2, p); in logf()
|
H A D | v_sinf.c | 51 n = v_fma_f32 (InvPi, r, Shift); in sinf() 56 r = v_fma_f32 (-Pi1, n, r); in sinf() 57 r = v_fma_f32 (-Pi2, n, r); in sinf() 58 r = v_fma_f32 (-Pi3, n, r); in sinf() 62 y = v_fma_f32 (A9, r2, A7); in sinf() 63 y = v_fma_f32 (y, r2, A5); in sinf() 64 y = v_fma_f32 (y, r2, A3); in sinf() 65 y = v_fma_f32 (y * r2, r, r); in sinf()
|
H A D | v_exp2f.c | 38 v_u32_t r1 = v_as_u32_f32 (v_fma_f32 (poly, s2, s2) * s1); in specialcase() 40 v_u32_t r0 = v_as_u32_f32 (v_fma_f32 (poly, scale, scale)); in specialcase() 68 p = v_fma_f32 (C0, r, C1); in exp2f() 69 q = v_fma_f32 (C2, r, C3); in exp2f() 70 q = v_fma_f32 (p, r2, q); in exp2f() 72 poly = v_fma_f32 (q, r2, p); in exp2f() 75 return v_fma_f32 (poly, scale, scale); in exp2f()
|
H A D | v_exp2f_1u.c | 65 poly = v_fma_f32 (C0, r, C1); in exp2f_1u() 66 poly = v_fma_f32 (poly, r, C2); in exp2f_1u() 67 poly = v_fma_f32 (poly, r, C3); in exp2f_1u() 68 poly = v_fma_f32 (poly, r, C4); in exp2f_1u() 69 poly = v_fma_f32 (poly, r, C5); in exp2f_1u() 70 poly = v_fma_f32 (poly, r, v_f32 (1.0f)); in exp2f_1u()
|
H A D | v_math.h | 180 v_fma_f32 (v_f32_t x, v_f32_t y, v_f32_t z) in v_fma_f32() function 449 v_fma_f32 (v_f32_t x, v_f32_t y, v_f32_t z) in v_fma_f32() function
|
/third_party/mesa3d/src/amd/compiler/ |
H A D | aco_optimizer.cpp | 3610 case aco_opcode::v_fma_f32: break; in can_use_mad_mix() 3616 if (instr->opcode == aco_opcode::v_fma_f32 && !ctx.program->dev.fused_mad_mix && in can_use_mad_mix() 3629 bool is_add = instr->opcode != aco_opcode::v_mul_f32 && instr->opcode != aco_opcode::v_fma_f32; in to_mad_mix() 3732 instr->opcode != aco_opcode::v_mul_f32 && instr->opcode != aco_opcode::v_fma_f32; in combine_mad_mix() 4085 aco_opcode mad_op = emit_fma ? aco_opcode::v_fma_f32 : aco_opcode::v_mad_f32; in combine_instruction() 4218 } else if (instr->opcode == aco_opcode::v_fma_f32 || instr->opcode == aco_opcode::v_fma_f16) { in combine_instruction() 4219 /* set existing v_fma_f32 with label_mad so we can create v_fmamk_f32/v_fmaak_f32. in combine_instruction() 4384 if ((instr->opcode == aco_opcode::v_fma_f32 || instr->opcode == aco_opcode::v_fma_f16) && in select_instruction() 4734 if (instr->opcode == aco_opcode::v_fma_f32) in apply_literals()
|
H A D | aco_register_allocation.cpp | 2441 case aco_opcode::v_fma_f32: in get_affinities() 2550 (instr->opcode != aco_opcode::v_fma_f32 || program->gfx_level < GFX10) && in optimize_encoding_vop2() 2583 case aco_opcode::v_fma_f32: instr->opcode = aco_opcode::v_fmac_f32; break; in optimize_encoding_vop2()
|
H A D | aco_instruction_selection.cpp | 2404 emit_vop3a_instruction(ctx, instr, aco_opcode::v_fma_f32, dst, 8211 ctx->program->gfx_level >= GFX10_3 ? aco_opcode::v_fma_f32 : aco_opcode::v_mad_f32;
|
/third_party/mesa3d/src/amd/compiler/tests/ |
H A D | helpers.cpp | 328 return b.vop3(aco_opcode::v_fma_f32, b.def(v1), src0, src1, src2); in fma()
|