radv: handle GFX9 1D textures

As GFX9 can't handle 1D depth textures, radeonsi and
apparantly pro just update all 1D textures to 2D,
and work around it.

This ports the workarounds from radeonsi.

Reviewed-by: Bas Nieuwenhuizen <[email protected]>
Cc: "17.2" <[email protected]>
Signed-off-by: Dave Airlie <[email protected]>

2 files changed, 76 insertions, 14 deletions

diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
index 8f9f771acf8..22e915dd0dd 100644
--- a/src/amd/common/ac_nir_to_llvm.c
+++ b/src/amd/common/ac_nir_to_llvm.c
@@ -3264,13 +3264,13 @@ static LLVMValueRef get_image_coords(struct ac_nir_context *ctx,
 
 	int count;
 	enum glsl_sampler_dim dim = glsl_get_sampler_dim(type);
+	bool is_array = glsl_sampler_type_is_array(type);
 	bool add_frag_pos = (dim == GLSL_SAMPLER_DIM_SUBPASS ||
 			     dim == GLSL_SAMPLER_DIM_SUBPASS_MS);
 	bool is_ms = (dim == GLSL_SAMPLER_DIM_MS ||
 		      dim == GLSL_SAMPLER_DIM_SUBPASS_MS);
-
-	count = image_type_to_components_count(dim,
-					       glsl_sampler_type_is_array(type));
+	bool gfx9_1d = ctx->abi->chip_class >= GFX9 && dim == GLSL_SAMPLER_DIM_1D;
+	count = image_type_to_components_count(dim, is_array);
 
 	if (is_ms) {
 		LLVMValueRef fmask_load_address[3];
@@ -3278,7 +3278,7 @@ static LLVMValueRef get_image_coords(struct ac_nir_context *ctx,
 
 		fmask_load_address[0] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[0], "");
 		fmask_load_address[1] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[1], "");
-		if (glsl_sampler_type_is_array(type))
+		if (is_array)
 			fmask_load_address[2] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[2], "");
 		else
 			fmask_load_address[2] = NULL;
@@ -3297,7 +3297,7 @@ static LLVMValueRef get_image_coords(struct ac_nir_context *ctx,
 							       sample_index,
 							       get_sampler_desc(ctx, instr->variables[0], AC_DESC_FMASK, true, false));
 	}
-	if (count == 1) {
+	if (count == 1 && !gfx9_1d) {
 		if (instr->src[0].ssa->num_components)
 			res = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[0], "");
 		else
@@ -3307,9 +3307,8 @@ static LLVMValueRef get_image_coords(struct ac_nir_context *ctx,
 		if (is_ms)
 			count--;
 		for (chan = 0; chan < count; ++chan) {
-			coords[chan] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[chan], "");
+			coords[chan] = llvm_extract_elem(&ctx->ac, src0, chan);
 		}
-
 		if (add_frag_pos) {
 			for (chan = 0; chan < 2; ++chan)
 				coords[chan] = LLVMBuildAdd(ctx->ac.builder, coords[chan], LLVMBuildFPToUI(ctx->ac.builder, ctx->abi->frag_pos[chan],
@@ -3317,6 +3316,16 @@ static LLVMValueRef get_image_coords(struct ac_nir_context *ctx,
 			coords[2] = ac_to_integer(&ctx->ac, ctx->abi->inputs[radeon_llvm_reg_index_soa(VARYING_SLOT_LAYER, 0)]);
 			count++;
 		}
+
+		if (gfx9_1d) {
+			if (is_array) {
+				coords[2] = coords[1];
+				coords[1] = ctx->ac.i32_0;
+			} else
+				coords[1] = ctx->ac.i32_0;
+			count++;
+		}
+
 		if (is_ms) {
 			coords[count] = sample_index;
 			count++;
@@ -3561,14 +3570,22 @@ static LLVMValueRef visit_image_size(struct ac_nir_context *ctx,
 
 	res = ac_build_image_opcode(&ctx->ac, &args);
 
+	LLVMValueRef two = LLVMConstInt(ctx->ac.i32, 2, false);
+
 	if (glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_CUBE &&
 	    glsl_sampler_type_is_array(type)) {
-		LLVMValueRef two = LLVMConstInt(ctx->ac.i32, 2, false);
 		LLVMValueRef six = LLVMConstInt(ctx->ac.i32, 6, false);
 		LLVMValueRef z = LLVMBuildExtractElement(ctx->ac.builder, res, two, "");
 		z = LLVMBuildSDiv(ctx->ac.builder, z, six, "");
 		res = LLVMBuildInsertElement(ctx->ac.builder, res, z, two, "");
 	}
+	if (glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_1D &&
+	    glsl_sampler_type_is_array(type)) {
+		LLVMValueRef layers = LLVMBuildExtractElement(ctx->ac.builder, res, two, "");
+		res = LLVMBuildInsertElement(ctx->ac.builder, res, layers,
+						ctx->ac.i32_1, "");
+
+	}
 	return res;
 }
 
@@ -4495,23 +4512,39 @@ static void visit_tex(struct ac_nir_context *ctx, nir_tex_instr *instr)
 
 	/* pack derivatives */
 	if (ddx || ddy) {
+		int num_src_deriv_channels, num_dest_deriv_channels;
 		switch (instr->sampler_dim) {
 		case GLSL_SAMPLER_DIM_3D:
 		case GLSL_SAMPLER_DIM_CUBE:
 			num_deriv_comp = 3;
+			num_src_deriv_channels = 3;
+			num_dest_deriv_channels = 3;
 			break;
 		case GLSL_SAMPLER_DIM_2D:
 		default:
+			num_src_deriv_channels = 2;
+			num_dest_deriv_channels = 2;
 			num_deriv_comp = 2;
 			break;
 		case GLSL_SAMPLER_DIM_1D:
-			num_deriv_comp = 1;
+			num_src_deriv_channels = 1;
+			if (ctx->abi->chip_class >= GFX9) {
+				num_dest_deriv_channels = 2;
+				num_deriv_comp = 2;
+			} else {
+				num_dest_deriv_channels = 1;
+				num_deriv_comp = 1;
+			}
 			break;
 		}
 
-		for (unsigned i = 0; i < num_deriv_comp; i++) {
+		for (unsigned i = 0; i < num_src_deriv_channels; i++) {
 			derivs[i] = ac_to_float(&ctx->ac, llvm_extract_elem(&ctx->ac, ddx, i));
-			derivs[num_deriv_comp + i] = ac_to_float(&ctx->ac, llvm_extract_elem(&ctx->ac, ddy, i));
+			derivs[num_dest_deriv_channels + i] = ac_to_float(&ctx->ac, llvm_extract_elem(&ctx->ac, ddy, i));
+		}
+		for (unsigned i = num_src_deriv_channels; i < num_dest_deriv_channels; i++) {
+			derivs[i] = ctx->ac.f32_0;
+			derivs[num_dest_deriv_channels + i] = ctx->ac.f32_0;
 		}
 	}
 
@@ -4552,6 +4585,23 @@ static void visit_tex(struct ac_nir_context *ctx, nir_tex_instr *instr)
 			}
 			address[count++] = coords[2];
 		}
+
+		if (ctx->abi->chip_class >= GFX9) {
+			LLVMValueRef filler;
+			if (instr->op == nir_texop_txf)
+				filler = ctx->ac.i32_0;
+			else
+				filler = LLVMConstReal(ctx->ac.f32, 0.5);
+
+			if (instr->sampler_dim == GLSL_SAMPLER_DIM_1D) {
+				if (instr->is_array) {
+					address[count] = address[count - 1];
+					address[count - 1] = filler;
+					count++;
+				} else
+					address[count++] = filler;
+			}
+		}
 	}
 
 	/* Pack LOD */
@@ -4648,6 +4698,14 @@ static void visit_tex(struct ac_nir_context *ctx, nir_tex_instr *instr)
 		LLVMValueRef z = LLVMBuildExtractElement(ctx->ac.builder, result, two, "");
 		z = LLVMBuildSDiv(ctx->ac.builder, z, six, "");
 		result = LLVMBuildInsertElement(ctx->ac.builder, result, z, two, "");
+	} else if (ctx->abi->chip_class >= GFX9 &&
+		   instr->op == nir_texop_txs &&
+		   instr->sampler_dim == GLSL_SAMPLER_DIM_1D &&
+		   instr->is_array) {
+		LLVMValueRef two = LLVMConstInt(ctx->ac.i32, 2, false);
+		LLVMValueRef layers = LLVMBuildExtractElement(ctx->ac.builder, result, two, "");
+		result = LLVMBuildInsertElement(ctx->ac.builder, result, layers,
+						ctx->ac.i32_1, "");
 	} else if (instr->dest.ssa.num_components != 4)
 		result = trim_vector(&ctx->ac, result, instr->dest.ssa.num_components);
 
diff --git a/src/amd/vulkan/radv_image.c b/src/amd/vulkan/radv_image.c
index 4d83ae564fc..df288666beb 100644
--- a/src/amd/vulkan/radv_image.c
+++ b/src/amd/vulkan/radv_image.c
@@ -281,10 +281,14 @@ si_set_mutable_tex_desc_fields(struct radv_device *device,
 }
 
 static unsigned radv_tex_dim(VkImageType image_type, VkImageViewType view_type,
-			     unsigned nr_layers, unsigned nr_samples, bool is_storage_image)
+			     unsigned nr_layers, unsigned nr_samples, bool is_storage_image, bool gfx9)
 {
 	if (view_type == VK_IMAGE_VIEW_TYPE_CUBE || view_type == VK_IMAGE_VIEW_TYPE_CUBE_ARRAY)
 		return is_storage_image ? V_008F1C_SQ_RSRC_IMG_2D_ARRAY : V_008F1C_SQ_RSRC_IMG_CUBE;
+
+	/* GFX9 allocates 1D textures as 2D. */
+	if (gfx9 && image_type == VK_IMAGE_TYPE_1D)
+		image_type = VK_IMAGE_TYPE_2D;
 	switch (image_type) {
 	case VK_IMAGE_TYPE_1D:
 		return nr_layers > 1 ? V_008F1C_SQ_RSRC_IMG_1D_ARRAY : V_008F1C_SQ_RSRC_IMG_1D;
@@ -375,7 +379,7 @@ si_make_texture_descriptor(struct radv_device *device,
 	}
 
 	type = radv_tex_dim(image->type, view_type, image->info.array_size, image->info.samples,
-			    is_storage_image);
+			    is_storage_image, device->physical_device->rad_info.chip_class >= GFX9);
 	if (type == V_008F1C_SQ_RSRC_IMG_1D_ARRAY) {
 	        height = 1;
 		depth = image->info.array_size;
@@ -495,7 +499,7 @@ si_make_texture_descriptor(struct radv_device *device,
 			S_008F1C_DST_SEL_Y(V_008F1C_SQ_SEL_X) |
 			S_008F1C_DST_SEL_Z(V_008F1C_SQ_SEL_X) |
 			S_008F1C_DST_SEL_W(V_008F1C_SQ_SEL_X) |
-			S_008F1C_TYPE(radv_tex_dim(image->type, view_type, 1, 0, false));
+			S_008F1C_TYPE(radv_tex_dim(image->type, view_type, 1, 0, false, false));
 		fmask_state[4] = 0;
 		fmask_state[5] = S_008F24_BASE_ARRAY(first_layer);
 		fmask_state[6] = 0;