ac/nir: Rewrite gather4 integer workaround based on radeonsi

The workaround was originally written based on amdgpu-pro traces, but
since then radeonsi has got its own slightly different version. Use the
radeonsi version instead, to be consistent and because it'll be slightly
more convenient for handling unnormalized coordinates.

Acked-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

1 files changed, 85 insertions, 65 deletions

diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
index 5acac64665b..5d6531558aa 100644
--- a/src/amd/common/ac_nir_to_llvm.c
+++ b/src/amd/common/ac_nir_to_llvm.c
@@ -1182,6 +1182,21 @@ get_buffer_size(struct ac_nir_context *ctx, LLVMValueRef descriptor, bool in_ele
 	return size;
 }
 
+/* Gather4 should follow the same rules as bilinear filtering, but the hardware
+ * incorrectly forces nearest filtering if the texture format is integer.
+ * The only effect it has on Gather4, which always returns 4 texels for
+ * bilinear filtering, is that the final coordinates are off by 0.5 of
+ * the texel size.
+ *
+ * The workaround is to subtract 0.5 from the unnormalized coordinates,
+ * or (0.5 / size) from the normalized coordinates.
+ *
+ * However, cube textures with 8_8_8_8 data formats require a different
+ * workaround of overriding the num format to USCALED/SSCALED. This would lose
+ * precision in 32-bit data formats, so it needs to be applied dynamically at
+ * runtime. In this case, return an i1 value that indicates whether the
+ * descriptor was overridden (and hence a fixup of the sampler result is needed).
+ */
 static LLVMValueRef lower_gather4_integer(struct ac_llvm_context *ctx,
 					  nir_variable *var,
 					  struct ac_image_args *args,
@@ -1189,33 +1204,87 @@ static LLVMValueRef lower_gather4_integer(struct ac_llvm_context *ctx,
 {
 	const struct glsl_type *type = glsl_without_array(var->type);
 	enum glsl_base_type stype = glsl_get_sampler_result_type(type);
+	LLVMValueRef wa_8888 = NULL;
 	LLVMValueRef half_texel[2];
-	LLVMValueRef compare_cube_wa = NULL;
 	LLVMValueRef result;
 
-	//TODO Rect
-	{
-		struct ac_image_args txq_args = { 0 };
+	assert(stype == GLSL_TYPE_INT || stype == GLSL_TYPE_UINT);
+
+	if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE) {
+		LLVMValueRef formats;
+		LLVMValueRef data_format;
+		LLVMValueRef wa_formats;
+
+		formats = LLVMBuildExtractElement(ctx->builder, args->resource, ctx->i32_1, "");
+
+		data_format = LLVMBuildLShr(ctx->builder, formats,
+					    LLVMConstInt(ctx->i32, 20, false), "");
+		data_format = LLVMBuildAnd(ctx->builder, data_format,
+					   LLVMConstInt(ctx->i32, (1u << 6) - 1, false), "");
+		wa_8888 = LLVMBuildICmp(
+			ctx->builder, LLVMIntEQ, data_format,
+			LLVMConstInt(ctx->i32, V_008F14_IMG_DATA_FORMAT_8_8_8_8, false),
+			"");
+
+		uint32_t wa_num_format =
+			stype == GLSL_TYPE_UINT ?
+			S_008F14_NUM_FORMAT(V_008F14_IMG_NUM_FORMAT_USCALED) :
+			S_008F14_NUM_FORMAT(V_008F14_IMG_NUM_FORMAT_SSCALED);
+		wa_formats = LLVMBuildAnd(ctx->builder, formats,
+					  LLVMConstInt(ctx->i32, C_008F14_NUM_FORMAT, false),
+					  "");
+		wa_formats = LLVMBuildOr(ctx->builder, wa_formats,
+					LLVMConstInt(ctx->i32, wa_num_format, false), "");
+
+		formats = LLVMBuildSelect(ctx->builder, wa_8888, wa_formats, formats, "");
+		args->resource = LLVMBuildInsertElement(
+			ctx->builder, args->resource, formats, ctx->i32_1, "");
+	}
+
+	if (instr->sampler_dim == GLSL_SAMPLER_DIM_RECT) {
+		assert(!wa_8888);
+		half_texel[0] = half_texel[1] = LLVMConstReal(ctx->f32, -0.5);
+	} else {
+		struct ac_image_args resinfo = {};
+		LLVMBasicBlockRef bbs[2];
+
+		bbs[0] = LLVMGetInsertBlock(ctx->builder);
+		if (wa_8888) {
+			/* Skip the texture size query entirely if we don't need it. */
+			ac_build_ifcc(ctx, LLVMBuildNot(ctx->builder, wa_8888, ""), 2000);
+			bbs[1] = LLVMGetInsertBlock(ctx->builder);
+		}
 
-		txq_args.dim = get_ac_sampler_dim(ctx, instr->sampler_dim, instr->is_array);
-		txq_args.opcode = ac_image_get_resinfo;
-		txq_args.dmask = 0xf;
-		txq_args.lod = ctx->i32_0;
-		txq_args.resource = args->resource;
-		txq_args.attributes = AC_FUNC_ATTR_READNONE;
-		LLVMValueRef size = ac_build_image_opcode(ctx, &txq_args);
+		/* Query the texture size. */
+		resinfo.dim = get_ac_sampler_dim(ctx, instr->sampler_dim, instr->is_array);
+		resinfo.opcode = ac_image_get_resinfo;
+		resinfo.dmask = 0xf;
+		resinfo.lod = ctx->i32_0;
+		resinfo.resource = args->resource;
+		resinfo.attributes = AC_FUNC_ATTR_READNONE;
+		LLVMValueRef size = ac_build_image_opcode(ctx, &resinfo);
 
+		/* Compute -0.5 / size. */
 		for (unsigned c = 0; c < 2; c++) {
-			half_texel[c] = LLVMBuildExtractElement(ctx->builder, size,
-								LLVMConstInt(ctx->i32, c, false), "");
+			half_texel[c] =
+				LLVMBuildExtractElement(ctx->builder, size,
+							LLVMConstInt(ctx->i32, c, 0), "");
 			half_texel[c] = LLVMBuildUIToFP(ctx->builder, half_texel[c], ctx->f32, "");
 			half_texel[c] = ac_build_fdiv(ctx, ctx->f32_1, half_texel[c]);
 			half_texel[c] = LLVMBuildFMul(ctx->builder, half_texel[c],
 						      LLVMConstReal(ctx->f32, -0.5), "");
 		}
-	}
 
-	LLVMValueRef orig_coords[2] = { args->coords[0], args->coords[1] };
+		if (wa_8888) {
+			ac_build_endif(ctx, 2000);
+
+			for (unsigned c = 0; c < 2; c++) {
+				LLVMValueRef values[2] = { ctx->f32_0, half_texel[c] };
+				half_texel[c] = ac_build_phi(ctx, ctx->f32, 2,
+							     values, bbs);
+			}
+		}
+	}
 
 	for (unsigned c = 0; c < 2; c++) {
 		LLVMValueRef tmp;
@@ -1223,55 +1292,6 @@ static LLVMValueRef lower_gather4_integer(struct ac_llvm_context *ctx,
 		args->coords[c] = LLVMBuildFAdd(ctx->builder, tmp, half_texel[c], "");
 	}
 
-	/*
-	 * Apparantly cube has issue with integer types that the workaround doesn't solve,
-	 * so this tests if the format is 8_8_8_8 and an integer type do an alternate
-	 * workaround by sampling using a scaled type and converting.
-	 * This is taken from amdgpu-pro shaders.
-	 */
-	/* NOTE this produces some ugly code compared to amdgpu-pro,
-	 * LLVM ends up dumping SGPRs into VGPRs to deal with the compare/select,
-	 * and then reads them back. -pro generates two selects,
-	 * one s_cmp for the descriptor rewriting
-	 * one v_cmp for the coordinate and result changes.
-	 */
-	if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE) {
-		LLVMValueRef tmp, tmp2;
-
-		/* workaround 8/8/8/8 uint/sint cube gather bug */
-		/* first detect it then change to a scaled read and f2i */
-		tmp = LLVMBuildExtractElement(ctx->builder, args->resource, ctx->i32_1, "");
-		tmp2 = tmp;
-
-		/* extract the DATA_FORMAT */
-		tmp = ac_build_bfe(ctx, tmp, LLVMConstInt(ctx->i32, 20, false),
-				   LLVMConstInt(ctx->i32, 6, false), false);
-
-		/* is the DATA_FORMAT == 8_8_8_8 */
-		compare_cube_wa = LLVMBuildICmp(ctx->builder, LLVMIntEQ, tmp, LLVMConstInt(ctx->i32, V_008F14_IMG_DATA_FORMAT_8_8_8_8, false), "");
-
-		if (stype == GLSL_TYPE_UINT)
-			/* Create a NUM FORMAT - 0x2 or 0x4 - USCALED or UINT */
-			tmp = LLVMBuildSelect(ctx->builder, compare_cube_wa, LLVMConstInt(ctx->i32, 0x8000000, false),
-					      LLVMConstInt(ctx->i32, 0x10000000, false), "");
-		else
-			/* Create a NUM FORMAT - 0x3 or 0x5 - SSCALED or SINT */
-			tmp = LLVMBuildSelect(ctx->builder, compare_cube_wa, LLVMConstInt(ctx->i32, 0xc000000, false),
-					      LLVMConstInt(ctx->i32, 0x14000000, false), "");
-
-		/* replace the NUM FORMAT in the descriptor */
-		tmp2 = LLVMBuildAnd(ctx->builder, tmp2, LLVMConstInt(ctx->i32, C_008F14_NUM_FORMAT, false), "");
-		tmp2 = LLVMBuildOr(ctx->builder, tmp2, tmp, "");
-
-		args->resource = LLVMBuildInsertElement(ctx->builder, args->resource, tmp2, ctx->i32_1, "");
-
-		/* don't modify the coordinates for this case */
-		for (unsigned c = 0; c < 2; ++c)
-			args->coords[c] = LLVMBuildSelect(
-				ctx->builder, compare_cube_wa,
-				orig_coords[c], args->coords[c], "");
-	}
-
 	args->attributes = AC_FUNC_ATTR_READNONE;
 	result = ac_build_image_opcode(ctx, args);
 
@@ -1287,7 +1307,7 @@ static LLVMValueRef lower_gather4_integer(struct ac_llvm_context *ctx,
 				tmp2 = LLVMBuildFPToSI(ctx->builder, tmp, ctx->i32, "");
 			tmp = LLVMBuildBitCast(ctx->builder, tmp, ctx->i32, "");
 			tmp2 = LLVMBuildBitCast(ctx->builder, tmp2, ctx->i32, "");
-			tmp = LLVMBuildSelect(ctx->builder, compare_cube_wa, tmp2, tmp, "");
+			tmp = LLVMBuildSelect(ctx->builder, wa_8888, tmp2, tmp, "");
 			tmp = LLVMBuildBitCast(ctx->builder, tmp, ctx->f32, "");
 			result = LLVMBuildInsertElement(ctx->builder, result, tmp, LLVMConstInt(ctx->i32, c, false), "");
 		}