freedreno/ir3: add a pass to lower tg4 to txl, enable gather on a4xx

Unfortunately Adreno A4xx hardware returns incorrect results with the
GATHER4 opcodes. As a result, we have to lower to 4 individual texture
calls (txl since we have to force lod to 0). We achieve this using
offsets, including on cube maps which normally never have offsets.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Rob Clark <robdclark@gmail.com>

docs/features.txt

@@ -130,7 +130,7 @@ GL 4.0, GLSL 4.00 --- all DONE: i965/gen7+, nvc0, r600, radeonsi
   GL_ARB_tessellation_shader                            DONE (i965/gen7+)
   GL_ARB_texture_buffer_object_rgb32                    DONE (freedreno, i965/gen6+, llvmpipe, softpipe, swr)
   GL_ARB_texture_cube_map_array                         DONE (i965/gen6+, nv50, llvmpipe, softpipe)
-  GL_ARB_texture_gather                                 DONE (freedreno/a5xx, i965/gen6+, nv50, llvmpipe, softpipe, swr)
+  GL_ARB_texture_gather                                 DONE (freedreno, i965/gen6+, nv50, llvmpipe, softpipe, swr)
   GL_ARB_texture_query_lod                              DONE (freedreno, i965, nv50, llvmpipe, softpipe)
   GL_ARB_transform_feedback2                            DONE (i965/gen6+, nv50, llvmpipe, softpipe, swr)
   GL_ARB_transform_feedback3                            DONE (i965/gen7+, llvmpipe, softpipe, swr)
@@ -256,7 +256,7 @@ GLES3.1, GLSL ES 3.1 -- all DONE: i965/hsw+, nvc0, radeonsi
   GL_ARB_texture_multisample (Multisample textures)     DONE (i965/gen7+, nv50, r600, llvmpipe, softpipe)
   GL_ARB_texture_storage_multisample                    DONE (all drivers that support GL_ARB_texture_multisample)
   GL_ARB_vertex_attrib_binding                          DONE (all drivers)
-  GS5 Enhanced textureGather                            DONE (i965/gen7+, r600)
+  GS5 Enhanced textureGather                            DONE (freedreno, i965/gen7+, r600)
   GS5 Packing/bitfield/conversion functions             DONE (i965/gen6+, r600)
   GL_EXT_shader_integer_mix                             DONE (all drivers that support GLSL)
 

src/gallium/drivers/freedreno/Makefile.sources

@@ -168,6 +168,7 @@ ir3_SOURCES := \
 	ir3/ir3_nir.c \
 	ir3/ir3_nir.h \
 	ir3/ir3_nir_lower_if_else.c \
+	ir3/ir3_nir_lower_tg4_to_tex.c \
 	ir3/ir3_print.c \
 	ir3/ir3_ra.c \
 	ir3/ir3_sched.c \

src/gallium/drivers/freedreno/freedreno_screen.c

@@ -264,7 +264,7 @@ fd_screen_get_param(struct pipe_screen *pscreen, enum pipe_cap param)
 		return 0;
 
 	case PIPE_CAP_MAX_TEXTURE_GATHER_COMPONENTS:
-		if (is_a5xx(screen))
+		if (is_a4xx(screen) || is_a5xx(screen))
 			return 4;
 		return 0;
 

src/gallium/drivers/freedreno/ir3/ir3_compiler_nir.c

@@ -2399,9 +2399,12 @@ emit_tex(struct ir3_context *ctx, nir_tex_instr *tex)
 	 */
 	if (has_off | has_lod | has_bias) {
 		if (has_off) {
-			for (i = 0; i < coords; i++)
+			unsigned off_coords = coords;
+			if (tex->sampler_dim == GLSL_SAMPLER_DIM_CUBE)
+				off_coords--;
+			for (i = 0; i < off_coords; i++)
 				src1[nsrc1++] = off[i];
-			if (coords < 2)
+			if (off_coords < 2)
 				src1[nsrc1++] = create_immed(b, fui(0.0));
 			flags |= IR3_INSTR_O;
 		}

src/gallium/drivers/freedreno/ir3/ir3_nir.c

@@ -188,6 +188,8 @@ ir3_optimize_nir(struct ir3_shader *shader, nir_shader *s,
 
 	OPT_V(s, nir_lower_tex, &tex_options);
 	OPT_V(s, nir_lower_load_const_to_scalar);
+	if (shader->compiler->gpu_id < 500)
+		OPT_V(s, ir3_nir_lower_tg4_to_tex);
 
 	ir3_optimize_loop(s);
 

src/gallium/drivers/freedreno/ir3/ir3_nir.h

@@ -38,6 +38,7 @@ void ir3_nir_scan_driver_consts(nir_shader *shader, struct ir3_driver_const_layo
 
 bool ir3_nir_lower_if_else(nir_shader *shader);
 bool ir3_nir_apply_trig_workarounds(nir_shader *shader);
+bool ir3_nir_lower_tg4_to_tex(nir_shader *shader);
 
 struct nir_shader * ir3_tgsi_to_nir(const struct tgsi_token *tokens);
 const nir_shader_compiler_options * ir3_get_compiler_options(struct ir3_compiler *compiler);

src/gallium/drivers/freedreno/ir3/ir3_nir_lower_tg4_to_tex.c

@@ -0,0 +1,140 @@
+/*
+ * Copyright © 2017 Ilia Mirkin
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include "ir3_nir.h"
+#include "compiler/nir/nir_builder.h"
+
+/* A4XX has a broken GATHER4 operation. It performs the texture swizzle on the
+ * gather results, rather than before. As a result, it must be emulated with
+ * direct texture calls.
+ */
+
+static bool
+lower_tg4(nir_block *block, nir_builder *b, void *mem_ctx)
+{
+	bool progress = false;
+
+	static const int offsets[3][2] = { {0, 1}, {1, 1}, {1, 0} };
+
+	nir_foreach_instr_safe(instr, block) {
+		if (instr->type != nir_instr_type_tex)
+			continue;
+
+        nir_tex_instr *tg4 = (nir_tex_instr *)instr;
+
+		if (tg4->op != nir_texop_tg4)
+			continue;
+
+		b->cursor = nir_before_instr(&tg4->instr);
+
+		nir_ssa_def *results[4];
+		int offset_index = nir_tex_instr_src_index(tg4, nir_tex_src_offset);
+		for (int i = 0; i < 4; i++) {
+			int num_srcs = tg4->num_srcs + 1 /* lod */;
+			if (offset_index < 0 && i < 3)
+				num_srcs++;
+
+			nir_tex_instr *tex = nir_tex_instr_create(b->shader, num_srcs);
+			tex->op = nir_texop_txl;
+			tex->sampler_dim = tg4->sampler_dim;
+			tex->coord_components = tg4->coord_components;
+			tex->is_array = tg4->is_array;
+			tex->is_shadow = tg4->is_shadow;
+			tex->is_new_style_shadow = tg4->is_new_style_shadow;
+			tex->texture_index = tg4->texture_index;
+			tex->texture = nir_deref_var_clone(tg4->texture, tex);
+			tex->sampler_index = tg4->sampler_index;
+			tex->sampler = nir_deref_var_clone(tg4->sampler, tex);
+			tex->dest_type = tg4->dest_type;
+
+			for (int j = 0; j < tg4->num_srcs; j++) {
+				nir_src_copy(&tex->src[j].src, &tg4->src[j].src, tex);
+				tex->src[j].src_type = tg4->src[j].src_type;
+			}
+			if (i != 3) {
+				nir_ssa_def *offset =
+					nir_vec2(b, nir_imm_int(b, offsets[i][0]),
+							 nir_imm_int(b, offsets[i][1]));
+				if (offset_index < 0) {
+					tex->src[tg4->num_srcs].src = nir_src_for_ssa(offset);
+					tex->src[tg4->num_srcs].src_type = nir_tex_src_offset;
+				} else {
+					assert(nir_tex_instr_src_size(tex, offset_index) == 2);
+					nir_ssa_def *orig = nir_ssa_for_src(
+							b, tex->src[offset_index].src, 2);
+					tex->src[offset_index].src =
+						nir_src_for_ssa(nir_iadd(b, orig, offset));
+				}
+			}
+			tex->src[num_srcs - 1].src = nir_src_for_ssa(nir_imm_float(b, 0));
+			tex->src[num_srcs - 1].src_type = nir_tex_src_lod;
+
+			nir_ssa_dest_init(&tex->instr, &tex->dest,
+							  nir_tex_instr_dest_size(tex), 32, NULL);
+			nir_builder_instr_insert(b, &tex->instr);
+
+			results[i] = nir_channel(b, &tex->dest.ssa, tg4->component);
+		}
+
+		nir_ssa_def *result = nir_vec4(b, results[0], results[1], results[2], results[3]);
+		nir_ssa_def_rewrite_uses(&tg4->dest.ssa, nir_src_for_ssa(result));
+
+		nir_instr_remove(&tg4->instr);
+
+		progress = true;
+	}
+
+	return progress;
+}
+
+static bool
+lower_tg4_func(nir_function_impl *impl)
+{
+	void *mem_ctx = ralloc_parent(impl);
+	nir_builder b;
+	nir_builder_init(&b, impl);
+
+	bool progress = false;
+	nir_foreach_block_safe(block, impl) {
+		progress |= lower_tg4(block, &b, mem_ctx);
+	}
+
+	if (progress)
+		nir_metadata_preserve(impl, nir_metadata_block_index |
+									nir_metadata_dominance);
+
+	return progress;
+}
+
+bool
+ir3_nir_lower_tg4_to_tex(nir_shader *shader)
+{
+	bool progress = false;
+
+	nir_foreach_function(function, shader) {
+		if (function->impl)
+			progress |= lower_tg4_func(function->impl);
+	}
+
+	return progress;
+}

src/gallium/drivers/freedreno/meson.build

@@ -188,6 +188,7 @@ files_libfreedreno = files(
   'ir3/ir3_nir.c',
   'ir3/ir3_nir.h',
   'ir3/ir3_nir_lower_if_else.c',
+  'ir3/ir3_nir_lower_tg4_to_tex.c',
   'ir3/ir3_print.c',
   'ir3/ir3_ra.c',
   'ir3/ir3_sched.c',