freedreno: a2xx: add partial lower_scalar pass for ir2
authorJonathan Marek <jonathan@marek.ca>
Thu, 13 Dec 2018 17:39:39 +0000 (12:39 -0500)
committerRob Clark <robdclark@gmail.com>
Tue, 22 Jan 2019 14:45:03 +0000 (14:45 +0000)
Some instructions can only be scalar on a2xx, lower these only

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
src/gallium/drivers/freedreno/Makefile.sources
src/gallium/drivers/freedreno/a2xx/ir2_nir.c
src/gallium/drivers/freedreno/a2xx/ir2_nir_lower_scalar.c [new file with mode: 0644]
src/gallium/drivers/freedreno/a2xx/ir2_private.h
src/gallium/drivers/freedreno/meson.build

index e576739b8abd94caad2f83987ad7e67ab0bade53..d2eddb989d732092707f2371b6f6bd9974727365 100644 (file)
@@ -70,6 +70,7 @@ a2xx_SOURCES := \
        a2xx/ir2_assemble.c \
        a2xx/ir2_cp.c \
        a2xx/ir2_nir.c \
+       a2xx/ir2_nir_lower_scalar.c \
        a2xx/ir2_private.h \
        a2xx/ir2_ra.c
 
index ef9c5e0c4df7a423aaf61b9d7ee7f8bcaa410397..8679116976b1fe37af2fe27509a32eab214114bc 100644 (file)
@@ -1106,6 +1106,9 @@ ir2_nir_compile(struct ir2_context *ctx, bool binning)
 
        OPT_V(ctx->nir, nir_lower_bool_to_float);
 
+       /* lower to scalar instructions that can only be scalar on a2xx */
+       OPT_V(ctx->nir, ir2_nir_lower_scalar);
+
        OPT_V(ctx->nir, nir_lower_locals_to_regs);
 
        OPT_V(ctx->nir, nir_convert_from_ssa, true);
diff --git a/src/gallium/drivers/freedreno/a2xx/ir2_nir_lower_scalar.c b/src/gallium/drivers/freedreno/a2xx/ir2_nir_lower_scalar.c
new file mode 100644 (file)
index 0000000..2b72a86
--- /dev/null
@@ -0,0 +1,174 @@
+/*
+ * Copyright (C) 2018 Jonathan Marek <jonathan@marek.ca>
+ *
+ * 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.
+ *
+ * Authors:
+ *    Jonathan Marek <jonathan@marek.ca>
+ */
+
+/* some operations can only be scalar on a2xx:
+ *  rsq, rcp, log2, exp2, cos, sin, sqrt
+ * mostly copy-pasted from nir_lower_alu_to_scalar.c
+ */
+
+#include "ir2_private.h"
+#include "compiler/nir/nir_builder.h"
+
+static void
+nir_alu_ssa_dest_init(nir_alu_instr * instr, unsigned num_components,
+                                         unsigned bit_size)
+{
+       nir_ssa_dest_init(&instr->instr, &instr->dest.dest, num_components,
+                                         bit_size, NULL);
+       instr->dest.write_mask = (1 << num_components) - 1;
+}
+
+static void
+lower_reduction(nir_alu_instr * instr, nir_op chan_op, nir_op merge_op,
+                               nir_builder * builder)
+{
+       unsigned num_components = nir_op_infos[instr->op].input_sizes[0];
+
+       nir_ssa_def *last = NULL;
+       for (unsigned i = 0; i < num_components; i++) {
+               nir_alu_instr *chan =
+                       nir_alu_instr_create(builder->shader, chan_op);
+               nir_alu_ssa_dest_init(chan, 1, instr->dest.dest.ssa.bit_size);
+               nir_alu_src_copy(&chan->src[0], &instr->src[0], chan);
+               chan->src[0].swizzle[0] = chan->src[0].swizzle[i];
+               if (nir_op_infos[chan_op].num_inputs > 1) {
+                       assert(nir_op_infos[chan_op].num_inputs == 2);
+                       nir_alu_src_copy(&chan->src[1], &instr->src[1], chan);
+                       chan->src[1].swizzle[0] = chan->src[1].swizzle[i];
+               }
+               chan->exact = instr->exact;
+
+               nir_builder_instr_insert(builder, &chan->instr);
+
+               if (i == 0) {
+                       last = &chan->dest.dest.ssa;
+               } else {
+                       last = nir_build_alu(builder, merge_op,
+                                                                last, &chan->dest.dest.ssa, NULL, NULL);
+               }
+       }
+
+       assert(instr->dest.write_mask == 1);
+       nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(last));
+       nir_instr_remove(&instr->instr);
+}
+
+static bool lower_scalar(nir_alu_instr * instr, nir_builder * b)
+{
+       assert(instr->dest.dest.is_ssa);
+       assert(instr->dest.write_mask != 0);
+
+       b->cursor = nir_before_instr(&instr->instr);
+       b->exact = instr->exact;
+
+#define LOWER_REDUCTION(name, chan, merge) \
+       case name##2: \
+       case name##3: \
+       case name##4: \
+               lower_reduction(instr, chan, merge, b); \
+               return true;
+
+       switch (instr->op) {
+               /* TODO: handle these instead of lowering */
+               LOWER_REDUCTION(nir_op_fall_equal, nir_op_seq, nir_op_fand);
+               LOWER_REDUCTION(nir_op_fany_nequal, nir_op_sne, nir_op_for);
+
+       default:
+               return false;
+       case nir_op_frsq:
+       case nir_op_frcp:
+       case nir_op_flog2:
+       case nir_op_fexp2:
+       case nir_op_fcos:
+       case nir_op_fsin:
+       case nir_op_fsqrt:
+               break;
+       }
+
+       assert(nir_op_infos[instr->op].num_inputs == 1);
+
+       unsigned num_components = instr->dest.dest.ssa.num_components;
+       nir_ssa_def *comps[NIR_MAX_VEC_COMPONENTS] = { NULL };
+       unsigned chan;
+
+       if (num_components == 1)
+               return false;
+
+       for (chan = 0; chan < num_components; chan++) {
+               assert(instr->dest.write_mask & (1 << chan));
+
+               nir_alu_instr *lower = nir_alu_instr_create(b->shader, instr->op);
+
+               nir_alu_src_copy(&lower->src[0], &instr->src[0], lower);
+               lower->src[0].swizzle[0] = instr->src[0].swizzle[chan];
+
+               nir_alu_ssa_dest_init(lower, 1, instr->dest.dest.ssa.bit_size);
+               lower->dest.saturate = instr->dest.saturate;
+               comps[chan] = &lower->dest.dest.ssa;
+               lower->exact = instr->exact;
+
+               nir_builder_instr_insert(b, &lower->instr);
+       }
+
+       nir_ssa_def *vec = nir_vec(b, comps, num_components);
+
+       nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(vec));
+
+       nir_instr_remove(&instr->instr);
+       return true;
+}
+
+static bool lower_scalar_impl(nir_function_impl * impl)
+{
+       nir_builder builder;
+       nir_builder_init(&builder, impl);
+       bool progress = false;
+
+       nir_foreach_block(block, impl) {
+               nir_foreach_instr_safe(instr, block) {
+                       if (instr->type == nir_instr_type_alu)
+                               progress = lower_scalar(nir_instr_as_alu(instr), &builder)
+                                       || progress;
+               }
+       }
+
+       nir_metadata_preserve(impl, nir_metadata_block_index |
+                                                 nir_metadata_dominance);
+
+       return progress;
+}
+
+bool ir2_nir_lower_scalar(nir_shader * shader)
+{
+       bool progress = false;
+
+       nir_foreach_function(function, shader) {
+               if (function->impl)
+                       progress = lower_scalar_impl(function->impl) || progress;
+       }
+
+       return progress;
+}
index 3bb3fe483de9b426c03148429d782fe8d6453a65..57c35ae810236ff06502fad84a50495ca52341b7 100644 (file)
@@ -188,6 +188,7 @@ struct ir2_context {
 void assemble(struct ir2_context *ctx, bool binning);
 
 void ir2_nir_compile(struct ir2_context *ctx, bool binning);
+bool ir2_nir_lower_scalar(nir_shader * shader);
 
 void ra_count_refs(struct ir2_context *ctx);
 void ra_reg(struct ir2_context *ctx, struct ir2_reg *reg, int force_idx,
index df1e617f60b74363f4547c65637a178fade72879..1e3a3037014bf59ceaa787bb7d1df4a31490ff48 100644 (file)
@@ -88,6 +88,7 @@ files_libfreedreno = files(
   'a2xx/ir2_assemble.c',
   'a2xx/ir2_cp.c',
   'a2xx/ir2_nir.c',
+  'a2xx/ir2_nir_lower_scalar.c',
   'a2xx/ir2_private.h',
   'a2xx/ir2_ra.c',
   'a3xx/fd3_blend.c',