OpenHarmony/third_party_mesa3d
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

ac/llvm: add option to clamp division by zero

Browse Source 
Replace div(x) by min(div(x), FLT_MAX)) to avoid getting a NaN result
when x is 0.

A cheaper alternative would be to use legacy mult instructions but they're
not exposed by LLVM.

Cc: mesa-stable
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6259>
This commit is contained in:
Pierre-Eric Pelloux-Prayer
2020-08-11 18:52:24 +02:00
parent f8c0e20152
commit 32f46a55c8
2 changed files with 10 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
src/amd/llvm/ac_nir_to_llvm.c
View File

@@ -713,6 +713,9 @@ static void visit_alu(struct ac_nir_context *ctx, const nir_alu_instr *instr)
result = emit_intrin_1f_param(&ctx->ac, "llvm.amdgcn.rcp", result = emit_intrin_1f_param(&ctx->ac, "llvm.amdgcn.rcp",
ac_to_float_type(&ctx->ac, def_type), src[0]); ac_to_float_type(&ctx->ac, def_type), src[0]);
} }
if (ctx->abi->clamp_div_by_zero)
result = ac_build_fmin(&ctx->ac, result,
LLVMConstReal(ac_to_float_type(&ctx->ac, def_type), FLT_MAX));
break; break;
case nir_op_iand: case nir_op_iand:
result = LLVMBuildAnd(ctx->ac.builder, src[0], src[1], ""); result = LLVMBuildAnd(ctx->ac.builder, src[0], src[1], "");
@@ -859,6 +862,9 @@ static void visit_alu(struct ac_nir_context *ctx, const nir_alu_instr *instr)
case nir_op_frsq: case nir_op_frsq:
result = emit_intrin_1f_param(&ctx->ac, "llvm.amdgcn.rsq", result = emit_intrin_1f_param(&ctx->ac, "llvm.amdgcn.rsq",
ac_to_float_type(&ctx->ac, def_type), src[0]); ac_to_float_type(&ctx->ac, def_type), src[0]);
if (ctx->abi->clamp_div_by_zero)
result = ac_build_fmin(&ctx->ac, result,
LLVMConstReal(ac_to_float_type(&ctx->ac, def_type), FLT_MAX));
break; break;
case nir_op_frexp_exp: case nir_op_frexp_exp:
src[0] = ac_to_float(&ctx->ac, src[0]); src[0] = ac_to_float(&ctx->ac, src[0]);
@@ -900,7 +906,7 @@ static void visit_alu(struct ac_nir_context *ctx, const nir_alu_instr *instr)
case nir_op_ffma: case nir_op_ffma:
/* FMA is better on GFX10, because it has FMA units instead of MUL-ADD units. */ /* FMA is better on GFX10, because it has FMA units instead of MUL-ADD units. */
result = emit_intrin_3f_param(&ctx->ac, ctx->ac.chip_class >= GFX10 ? "llvm.fma" : "llvm.fmuladd", result = emit_intrin_3f_param(&ctx->ac, ctx->ac.chip_class >= GFX10 ? "llvm.fma" : "llvm.fmuladd",
ac_to_float_type(&ctx->ac, def_type), src[0], src[1], src[2]); ac_to_float_type(&ctx->ac, def_type), src[0], src[1], src[2]);
break; break;
case nir_op_ldexp: case nir_op_ldexp:
src[0] = ac_to_float(&ctx->ac, src[0]); src[0] = ac_to_float(&ctx->ac, src[0]);

3
src/amd/llvm/ac_shader_abi.h
View File

@@ -192,6 +192,9 @@ struct ac_shader_abi {
/* Whether undef values must be converted to zero */ /* Whether undef values must be converted to zero */
bool convert_undef_to_zero; bool convert_undef_to_zero;
/* Clamp div by 0 (so it won't produce NaN) */
bool clamp_div_by_zero;
}; };
#endif /* AC_SHADER_ABI_H */ #endif /* AC_SHADER_ABI_H */