[Mesa-dev] [PATCH 2/2] radeonsi: fix a regression in integer cube map handling

[Nicolai Hähnle]

From: Nicolai Hähnle <nicolai.haehnle at amd.com>

A recent commit fixed the case of 8888 integer cube maps, which need the
workaround of replacing the data format with USCALED/SSCALED. However,
this broke the case of non-8888 integer cube maps; those still need the
fix of shifting the texture coordinates.

Fixes KHR-GL45.texture_gather.plain-gather-int-cube-array and similar.

Fixes: 6fb0c1013b35 ("radeonsi: workaround for gather4 on integer cube maps")
---
 src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c | 34 +++++++++++++++++------
 1 file changed, 26 insertions(+), 8 deletions(-)

diff --git a/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c b/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c
index be92044750c..0863876ed89 100644
--- a/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c
+++ b/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c
@@ -1715,72 +1715,78 @@ static void tex_fetch_args(
  * 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
 si_lower_gather4_integer(struct si_shader_context *ctx,
 			 struct ac_image_args *args,
 			 unsigned target,
 			 enum tgsi_return_type return_type)
 {
 	LLVMBuilderRef builder = ctx->gallivm.builder;
+	LLVMValueRef wa_8888 = NULL;
 	LLVMValueRef coord = args->addr;
 	LLVMValueRef half_texel[2];
 	/* Texture coordinates start after:
 	 *   {offset, bias, z-compare, derivatives}
 	 * Only the offset and z-compare can occur here.
 	 */
 	unsigned coord_vgpr_index = (int)args->offset + (int)args->compare;
 	int c;
 
 	assert(return_type == TGSI_RETURN_TYPE_SINT ||
 	       return_type == TGSI_RETURN_TYPE_UINT);
 
 	if (target == TGSI_TEXTURE_CUBE ||
 	    target == TGSI_TEXTURE_CUBE_ARRAY) {
 		LLVMValueRef formats;
 		LLVMValueRef data_format;
 		LLVMValueRef wa_formats;
-		LLVMValueRef wa;
 
 		formats = LLVMBuildExtractElement(builder, args->resource, ctx->i32_1, "");
 
 		data_format = LLVMBuildLShr(builder, formats,
 					    LLVMConstInt(ctx->i32, 20, false), "");
 		data_format = LLVMBuildAnd(builder, data_format,
 					   LLVMConstInt(ctx->i32, (1u << 6) - 1, false), "");
-		wa = LLVMBuildICmp(builder, LLVMIntEQ, data_format,
-				   LLVMConstInt(ctx->i32, V_008F14_IMG_DATA_FORMAT_8_8_8_8, false),
-				   "");
+		wa_8888 = LLVMBuildICmp(
+			builder, LLVMIntEQ, data_format,
+			LLVMConstInt(ctx->i32, V_008F14_IMG_DATA_FORMAT_8_8_8_8, false),
+			"");
 
 		uint32_t wa_num_format =
 			return_type == TGSI_RETURN_TYPE_UINT ?
 			S_008F14_NUM_FORMAT_GFX6(V_008F14_IMG_NUM_FORMAT_USCALED) :
 			S_008F14_NUM_FORMAT_GFX6(V_008F14_IMG_NUM_FORMAT_SSCALED);
 		wa_formats = LLVMBuildAnd(builder, formats,
 					  LLVMConstInt(ctx->i32, C_008F14_NUM_FORMAT_GFX6, false),
 					  "");
 		wa_formats = LLVMBuildOr(builder, wa_formats,
 					LLVMConstInt(ctx->i32, wa_num_format, false), "");
 
-		formats = LLVMBuildSelect(builder, wa, wa_formats, formats, "");
+		formats = LLVMBuildSelect(builder, wa_8888, wa_formats, formats, "");
 		args->resource = LLVMBuildInsertElement(
 			builder, args->resource, formats, ctx->i32_1, "");
-
-		return wa;
 	}
 
 	if (target == TGSI_TEXTURE_RECT ||
 	    target == TGSI_TEXTURE_SHADOWRECT) {
+		assert(!wa_8888);
 		half_texel[0] = half_texel[1] = LLVMConstReal(ctx->f32, -0.5);
 	} else {
 		struct tgsi_full_instruction txq_inst = {};
 		struct lp_build_emit_data txq_emit_data = {};
+		struct lp_build_if_state if_ctx;
+
+		if (wa_8888) {
+			/* Skip the texture size query entirely if we don't need it. */
+			lp_build_if(&if_ctx, &ctx->gallivm, LLVMBuildNot(builder, wa_8888, ""));
+		}
 
 		/* Query the texture size. */
 		txq_inst.Texture.Texture = target;
 		txq_emit_data.inst = &txq_inst;
 		txq_emit_data.dst_type = ctx->v4i32;
 		set_tex_fetch_args(ctx, &txq_emit_data, target,
 				   args->resource, NULL, &ctx->i32_0,
 				   1, 0xf);
 		txq_emit(NULL, &ctx->bld_base, &txq_emit_data);
 
@@ -1789,36 +1795,48 @@ si_lower_gather4_integer(struct si_shader_context *ctx,
 			half_texel[c] =
 				LLVMBuildExtractElement(builder, txq_emit_data.output[0],
 							LLVMConstInt(ctx->i32, c, 0), "");
 			half_texel[c] = LLVMBuildUIToFP(builder, half_texel[c], ctx->f32, "");
 			half_texel[c] =
 				lp_build_emit_llvm_unary(&ctx->bld_base,
 							 TGSI_OPCODE_RCP, half_texel[c]);
 			half_texel[c] = LLVMBuildFMul(builder, half_texel[c],
 						      LLVMConstReal(ctx->f32, -0.5), "");
 		}
+
+		if (wa_8888) {
+			lp_build_endif(&if_ctx);
+
+			LLVMBasicBlockRef bb[2] = { if_ctx.true_block, if_ctx.entry_block };
+
+			for (c = 0; c < 2; c++) {
+				LLVMValueRef values[2] = { half_texel[c], ctx->ac.f32_0 };
+				half_texel[c] = ac_build_phi(&ctx->ac, ctx->f32, 2,
+							     values, bb);
+			}
+		}
 	}
 
 	for (c = 0; c < 2; c++) {
 		LLVMValueRef tmp;
 		LLVMValueRef index = LLVMConstInt(ctx->i32, coord_vgpr_index + c, 0);
 
 		tmp = LLVMBuildExtractElement(builder, coord, index, "");
 		tmp = LLVMBuildBitCast(builder, tmp, ctx->f32, "");
 		tmp = LLVMBuildFAdd(builder, tmp, half_texel[c], "");
 		tmp = LLVMBuildBitCast(builder, tmp, ctx->i32, "");
 		coord = LLVMBuildInsertElement(builder, coord, tmp, index, "");
 	}
 
 	args->addr = coord;
 
-	return NULL;
+	return wa_8888;
 }
 
 /* The second half of the cube texture 8_8_8_8 integer workaround: adjust the
  * result after the gather operation.
  */
 static LLVMValueRef
 si_fix_gather4_integer_result(struct si_shader_context *ctx,
 			   LLVMValueRef result,
 			   enum tgsi_return_type return_type,
 			   LLVMValueRef wa)
-- 
2.11.0