ac: unify denorm setting enforcement

Acked-by: Pierre-Eric Pelloux-Prayer <[email protected]>
Reviewed-by: Samuel Pitoiset <[email protected]
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4196>

3 files changed, 14 insertions, 32 deletions

diff --git a/src/amd/common/ac_binary.c b/src/amd/common/ac_binary.c
index 5f92a57d7bf..8761422bd6b 100644
--- a/src/amd/common/ac_binary.c
+++ b/src/amd/common/ac_binary.c
@@ -58,11 +58,13 @@ void ac_parse_shader_binary_config(const char *data, size_t nbytes,
 				conf->num_vgprs = MAX2(conf->num_vgprs, (G_00B028_VGPRS(value) + 1) * 4);
 
 			conf->num_sgprs = MAX2(conf->num_sgprs, (G_00B028_SGPRS(value) + 1) * 8);
+			/* TODO: LLVM doesn't set FLOAT_MODE for non-compute shaders */
 			conf->float_mode =  G_00B028_FLOAT_MODE(value);
 			conf->rsrc1 = value;
 			break;
 		case R_00B02C_SPI_SHADER_PGM_RSRC2_PS:
 			conf->lds_size = MAX2(conf->lds_size, G_00B02C_EXTRA_LDS_SIZE(value));
+			/* TODO: LLVM doesn't set SHARED_VGPR_CNT for all shader types */
 			conf->num_shared_vgprs = G_00B02C_SHARED_VGPR_CNT(value);
 			conf->rsrc2 = value;
 			break;
@@ -124,4 +126,15 @@ void ac_parse_shader_binary_config(const char *data, size_t nbytes,
 		/* sgprs spills aren't spilling */
 	        conf->scratch_bytes_per_wave = G_00B860_WAVESIZE(scratch_size) * 256 * 4;
 	}
+
+	/* Enable 64-bit and 16-bit denormals, because there is no performance
+	 * cost.
+	 *
+	 * Don't enable denormals for 32-bit floats, because:
+	 * - denormals disable output modifiers
+	 * - denormals break v_mad_f32
+	 * - GFX6 & GFX7 would be very slow
+	 */
+	conf->float_mode &= ~V_00B028_FP_ALL_DENORMS;
+	conf->float_mode |= V_00B028_FP_64_DENORMS;
 }
diff --git a/src/amd/vulkan/radv_shader.c b/src/amd/vulkan/radv_shader.c
index 98c98db5665..70a51ee01d0 100644
--- a/src/amd/vulkan/radv_shader.c
+++ b/src/amd/vulkan/radv_shader.c
@@ -966,20 +966,6 @@ radv_shader_variant_create(struct radv_device *device,
 			return NULL;
 		}
 
-		/* Enable 64-bit and 16-bit denormals, because there is no performance
-		 * cost.
-		 *
-		 * If denormals are enabled, all floating-point output modifiers are
-		 * ignored.
-		 *
-		 * Don't enable denormals for 32-bit floats, because:
-		 * - Floating-point output modifiers would be ignored by the hw.
-		 * - Some opcodes don't support denormals, such as v_mad_f32. We would
-		 *   have to stop using those.
-		 * - GFX6 & GFX7 would be very slow.
-		 */
-		config.float_mode |= V_00B028_FP_64_DENORMS;
-
 		if (rtld_binary.lds_size > 0) {
 			unsigned alloc_granularity = device->physical_device->rad_info.chip_class >= GFX7 ? 512 : 256;
 			config.lds_size = align(rtld_binary.lds_size, alloc_granularity) / alloc_granularity;
diff --git a/src/gallium/drivers/radeonsi/si_shader_llvm.c b/src/gallium/drivers/radeonsi/si_shader_llvm.c
index 12a6d846c35..dca604afe40 100644
--- a/src/gallium/drivers/radeonsi/si_shader_llvm.c
+++ b/src/gallium/drivers/radeonsi/si_shader_llvm.c
@@ -130,24 +130,7 @@ bool si_compile_llvm(struct si_screen *sscreen,
 
 	bool ok = ac_rtld_read_config(&rtld, conf);
 	ac_rtld_close(&rtld);
-	if (!ok)
-		return false;
-
-	/* Enable 64-bit and 16-bit denormals, because there is no performance
-	 * cost.
-	 *
-	 * If denormals are enabled, all floating-point output modifiers are
-	 * ignored.
-	 *
-	 * Don't enable denormals for 32-bit floats, because:
-	 * - Floating-point output modifiers would be ignored by the hw.
-	 * - Some opcodes don't support denormals, such as v_mad_f32. We would
-	 *   have to stop using those.
-	 * - GFX6 & GFX7 would be very slow.
-	 */
-	conf->float_mode |= V_00B028_FP_64_DENORMS;
-
-	return true;
+	return ok;
 }
 
 void si_llvm_context_init(struct si_shader_context *ctx,