freedreno/ir3: Enable PIPE_CAP_PACKED_UNIFORMS

This commit turns on the gallium cap and adds a pass to lower the
load_ubo intrinsics for block 0 back to load_uniform intrinsics and
adjust the backend where the cap switches units from vec4s to dwords.

As we stop using ir3_glsl_type_size() for uniform layout, this also
corrects an issue where we would allocate a vec4 slot for samplers in
uniforms, fixing:

  dEQP-GLES3.functional.shaders.struct.uniform.sampler_array_fragment
  dEQP-GLES3.functional.shaders.struct.uniform.sampler_array_vertex
  dEQP-GLES3.functional.shaders.struct.uniform.sampler_nested_fragment
  dEQP-GLES2.functional.shaders.struct.uniform.sampler_nested_vertex
  dEQP-GLES2.functional.shaders.struct.uniform.sampler_nested_fragment

Signed-off-by: Kristian H. Kristensen <hoegsberg@chromium.org>
Reviewed-by: Rob Clark <robdclark@gmail.com>

src/freedreno/Makefile.sources

@@ -35,6 +35,7 @@ ir3_SOURCES := \
 	ir3/ir3_legalize.c \
 	ir3/ir3_nir.c \
 	ir3/ir3_nir.h \
+	ir3/ir3_nir_analyze_ubo_ranges.c \
 	ir3/ir3_nir_lower_io_offsets.c \
 	ir3/ir3_nir_lower_tg4_to_tex.c \
 	ir3/ir3_print.c \

src/freedreno/ir3/ir3_compiler_nir.c

@@ -681,8 +681,10 @@ emit_intrinsic_load_ubo(struct ir3_context *ctx, nir_intrinsic_instr *intr,
 	struct ir3_block *b = ctx->block;
 	struct ir3_instruction *base_lo, *base_hi, *addr, *src0, *src1;
 	nir_const_value *const_offset;
-	/* UBO addresses are the first driver params: */
-	unsigned ubo = regid(ctx->so->constbase.ubo, 0);
+	/* UBO addresses are the first driver params, but subtract 2 here to
+	 * account for nir_lower_uniforms_to_ubo rebasing the UBOs such that UBO 0
+	 * is the uniforms: */
+	unsigned ubo = regid(ctx->so->constbase.ubo, 0) - 2;
 	const unsigned ptrsz = ir3_pointer_size(ctx);
 
 	int off = 0;
@@ -1151,15 +1153,13 @@ emit_intrinsic(struct ir3_context *ctx, nir_intrinsic_instr *intr)
 		if (const_offset) {
 			idx += const_offset->u32[0];
 			for (int i = 0; i < intr->num_components; i++) {
-				unsigned n = idx * 4 + i;
-				dst[i] = create_uniform(b, n);
+				dst[i] = create_uniform(b, idx + i);
 			}
 		} else {
 			src = ir3_get_src(ctx, &intr->src[0]);
 			for (int i = 0; i < intr->num_components; i++) {
-				int n = idx * 4 + i;
-				dst[i] = create_uniform_indirect(b, n,
-						ir3_get_addr(ctx, src[0], 4));
+				dst[i] = create_uniform_indirect(b, idx + i,
+						ir3_get_addr(ctx, src[0], 1));
 			}
 			/* NOTE: if relative addressing is used, we set
 			 * constlen in the compiler (to worst-case value)

src/freedreno/ir3/ir3_nir.c

@@ -224,10 +224,13 @@ ir3_optimize_nir(struct ir3_shader *shader, nir_shader *s,
 
 	ir3_optimize_loop(s);
 
-	/* do idiv lowering after first opt loop to give a chance for
-	 * divide by immed power-of-two to be caught first:
+	/* do ubo load and idiv lowering after first opt loop to get a chance to
+	 * propagate constants for divide by immed power-of-two and constant ubo
+	 * block/offsets:
 	 */
-	if (OPT(s, nir_lower_idiv))
+	const bool ubo_progress = OPT(s, ir3_nir_analyze_ubo_ranges, shader);
+	const bool idiv_progress = OPT(s, nir_lower_idiv);
+	if (ubo_progress || idiv_progress)
 		ir3_optimize_loop(s);
 
 	OPT_V(s, nir_remove_dead_variables, nir_var_function_temp);

src/freedreno/ir3/ir3_nir.h

@@ -28,6 +28,7 @@
 #define IR3_NIR_H_
 
 #include "compiler/nir/nir.h"
+#include "compiler/nir/nir_builder.h"
 #include "compiler/shader_enums.h"
 
 #include "ir3_shader.h"
@@ -43,4 +44,9 @@ bool ir3_key_lowers_nir(const struct ir3_shader_key *key);
 struct nir_shader * ir3_optimize_nir(struct ir3_shader *shader, nir_shader *s,
 		const struct ir3_shader_key *key);
 
+bool ir3_nir_analyze_ubo_ranges(nir_shader *nir, struct ir3_shader *shader);
+
+nir_ssa_def *
+ir3_nir_try_propagate_bit_shift(nir_builder *b, nir_ssa_def *offset, int32_t shift);
+
 #endif /* IR3_NIR_H_ */

src/freedreno/ir3/ir3_nir_analyze_ubo_ranges.c

@@ -0,0 +1,95 @@
+/*
+ * Copyright © 2019 Google, Inc.
+ *
+ * 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.h"
+#include "compiler/nir/nir_builder.h"
+#include "util/u_dynarray.h"
+#include "mesa/main/macros.h"
+
+struct ir3_ubo_analysis_state {
+	unsigned lower_count;
+};
+
+static void
+lower_ubo_load_to_uniform(nir_intrinsic_instr *instr, nir_builder *b,
+						  struct ir3_ubo_analysis_state *state)
+{
+	/* We don't lower dynamic block index UBO loads to load_uniform, but we
+	 * could probably with some effort determine a block stride in number of
+	 * registers.
+	 */
+	if (!nir_src_is_const(instr->src[0]))
+		return;
+
+	const uint32_t block = nir_src_as_uint(instr->src[0]);
+	if (block > 0)
+		return;
+
+	b->cursor = nir_before_instr(&instr->instr);
+
+	nir_ssa_def *ubo_offset = nir_ssa_for_src(b, instr->src[1], 1);
+	nir_ssa_def *uniform_offset = ir3_nir_try_propagate_bit_shift(b, ubo_offset, -2);
+	if (uniform_offset == NULL)
+		uniform_offset = nir_ushr(b, ubo_offset, nir_imm_int(b, 2));
+
+	nir_intrinsic_instr *uniform =
+		nir_intrinsic_instr_create(b->shader, nir_intrinsic_load_uniform);
+	uniform->num_components = instr->num_components;
+	uniform->src[0] = nir_src_for_ssa(uniform_offset);
+	nir_ssa_dest_init(&uniform->instr, &uniform->dest,
+					  uniform->num_components, instr->dest.ssa.bit_size,
+					  instr->dest.ssa.name);
+	nir_builder_instr_insert(b, &uniform->instr);
+	nir_ssa_def_rewrite_uses(&instr->dest.ssa,
+							 nir_src_for_ssa(&uniform->dest.ssa));
+
+	nir_instr_remove(&instr->instr);
+
+	state->lower_count++;
+}
+
+bool
+ir3_nir_analyze_ubo_ranges(nir_shader *nir, struct ir3_shader *shader)
+{
+	struct ir3_ubo_analysis_state state = { 0 };
+
+	nir_foreach_function(function, nir) {
+		if (function->impl) {
+			nir_builder builder;
+			nir_builder_init(&builder, function->impl);
+			nir_foreach_block(block, function->impl) {
+				nir_foreach_instr_safe(instr, block) {
+					if (instr->type == nir_instr_type_intrinsic &&
+						nir_instr_as_intrinsic(instr)->intrinsic == nir_intrinsic_load_ubo)
+						lower_ubo_load_to_uniform(nir_instr_as_intrinsic(instr), &builder, &state);
+				}
+			}
+
+			nir_metadata_preserve(function->impl, nir_metadata_block_index |
+								  nir_metadata_dominance);
+		}
+	}
+
+	return state.lower_count > 0;
+}

src/freedreno/ir3/ir3_nir_lower_io_offsets.c

@@ -124,8 +124,8 @@ check_and_propagate_bit_shift32(nir_builder *b, nir_ssa_def *offset,
 	return shift_ssa;
 }
 
-static nir_ssa_def *
-try_propagate_bit_shift(nir_builder *b, nir_ssa_def *offset, int32_t shift)
+nir_ssa_def *
+ir3_nir_try_propagate_bit_shift(nir_builder *b, nir_ssa_def *offset, int32_t shift)
 {
 	nir_instr *offset_instr = offset->parent_instr;
 	if (offset_instr->type != nir_instr_type_alu)
@@ -187,7 +187,7 @@ lower_offset_for_ssbo(nir_intrinsic_instr *intrinsic, nir_builder *b,
 	 * Here we use the convention that shifting right is negative while shifting
 	 * left is positive. So 'x / 4' ~ 'x >> 2' or 'x << -2'.
 	 */
-	nir_ssa_def *new_offset = try_propagate_bit_shift(b, offset, -2);
+	nir_ssa_def *new_offset = ir3_nir_try_propagate_bit_shift(b, offset, -2);
 
 	/* The new source that will hold the dword-offset is always the last
 	 * one for every intrinsic.

src/freedreno/ir3/meson.build

@@ -50,6 +50,7 @@ libfreedreno_ir3_files = files(
   'ir3_legalize.c',
   'ir3_nir.c',
   'ir3_nir.h',
+  'ir3_nir_analyze_ubo_ranges.c',
   'ir3_nir_lower_io_offsets.c',
   'ir3_nir_lower_tg4_to_tex.c',
   'ir3_print.c',

src/gallium/drivers/freedreno/freedreno_screen.c

@@ -195,6 +195,7 @@ fd_screen_get_param(struct pipe_screen *pscreen, enum pipe_cap param)
 	case PIPE_CAP_MIXED_COLOR_DEPTH_BITS:
 	case PIPE_CAP_TEXTURE_BARRIER:
 	case PIPE_CAP_INVALIDATE_BUFFER:
+	case PIPE_CAP_PACKED_UNIFORMS:
 		return 1;
 
 	case PIPE_CAP_VERTEXID_NOBASE: