v3d: Use the new lower_to_scratch implementation for indirects on temps.
authorEric Anholt <eric@anholt.net>
Thu, 11 Apr 2019 18:12:01 +0000 (11:12 -0700)
committerEric Anholt <eric@anholt.net>
Fri, 12 Apr 2019 23:16:58 +0000 (16:16 -0700)
We can use the same register spilling infrastructure for our loads/stores
of indirect access of temp variables, instead of doing an if ladder.

Cuts 50% of instructions and max-temps from 2 KSP shaders in shader-db.
Also causes several other KSP shaders with large bodies and large loop
counts to not be force-unrolled.

The change was originally motivated by NOLTIS slightly modifying register
pressure in piglit temp mat4 array read/write tests, triggering register
allocation failures.

src/broadcom/Makefile.sources
src/broadcom/compiler/meson.build
src/broadcom/compiler/nir_to_vir.c
src/broadcom/compiler/v3d_compiler.h
src/broadcom/compiler/v3d_nir_lower_scratch.c [new file with mode: 0644]
src/broadcom/compiler/vir.c
src/broadcom/compiler/vir_register_allocate.c
src/gallium/drivers/v3d/v3d_screen.c

index 568e7b8701ba08c3cac29190ca7cd1a8ea46f50e..e759b545496d1dcf5860ed6c9cf43174d09fa17d 100644 (file)
@@ -41,6 +41,7 @@ BROADCOM_FILES = \
        compiler/v3d_compiler.h \
        compiler/v3d_nir_lower_image_load_store.c \
        compiler/v3d_nir_lower_io.c \
+       compiler/v3d_nir_lower_scratch.c \
        compiler/v3d_nir_lower_txf_ms.c \
        qpu/qpu_disasm.c \
        qpu/qpu_disasm.h \
index a58a711a83a1c76e509996c6d4a5e35a62fbdca0..1249af5e8ab2ce1dd0db1a8fc5b382290e54bef2 100644 (file)
@@ -37,6 +37,7 @@ libbroadcom_compiler_files = files(
   'v3d_compiler.h',
   'v3d_nir_lower_io.c',
   'v3d_nir_lower_image_load_store.c',
+  'v3d_nir_lower_scratch.c',
   'v3d_nir_lower_txf_ms.c',
 )
 
index de1ee07be7186c49f6a2132b3b4047d0d36753a6..11c49f5efa9302f2a8be59c32e647bac926d9719 100644 (file)
@@ -132,9 +132,11 @@ v3d_general_tmu_op(nir_intrinsic_instr *instr)
         case nir_intrinsic_load_ubo:
         case nir_intrinsic_load_uniform:
         case nir_intrinsic_load_shared:
+        case nir_intrinsic_load_scratch:
                 return GENERAL_TMU_READ_OP_READ;
         case nir_intrinsic_store_ssbo:
         case nir_intrinsic_store_shared:
+        case nir_intrinsic_store_scratch:
                 return GENERAL_TMU_WRITE_OP_WRITE;
         case nir_intrinsic_ssbo_atomic_add:
         case nir_intrinsic_shared_atomic_add:
@@ -177,7 +179,7 @@ v3d_general_tmu_op(nir_intrinsic_instr *instr)
  */
 static void
 ntq_emit_tmu_general(struct v3d_compile *c, nir_intrinsic_instr *instr,
-                     bool is_shared)
+                     bool is_shared_or_scratch)
 {
         /* XXX perf: We should turn add/sub of 1 to inc/dec.  Perhaps NIR
          * wants to have support for inc/dec?
@@ -185,8 +187,9 @@ ntq_emit_tmu_general(struct v3d_compile *c, nir_intrinsic_instr *instr,
 
         uint32_t tmu_op = v3d_general_tmu_op(instr);
         bool is_store = (instr->intrinsic == nir_intrinsic_store_ssbo ||
+                         instr->intrinsic == nir_intrinsic_store_scratch ||
                          instr->intrinsic == nir_intrinsic_store_shared);
-        bool has_index = !is_shared;
+        bool has_index = !is_shared_or_scratch;
 
         int offset_src;
         int tmu_writes = 1; /* address */
@@ -194,6 +197,7 @@ ntq_emit_tmu_general(struct v3d_compile *c, nir_intrinsic_instr *instr,
                 offset_src = 0;
         } else if (instr->intrinsic == nir_intrinsic_load_ssbo ||
                    instr->intrinsic == nir_intrinsic_load_ubo ||
+                   instr->intrinsic == nir_intrinsic_load_scratch ||
                    instr->intrinsic == nir_intrinsic_load_shared) {
                 offset_src = 0 + has_index;
         } else if (is_store) {
@@ -244,13 +248,18 @@ ntq_emit_tmu_general(struct v3d_compile *c, nir_intrinsic_instr *instr,
                 offset = vir_uniform(c, QUNIFORM_UBO_ADDR,
                                      v3d_unit_data_create(index, const_offset));
                 const_offset = 0;
-        } else if (is_shared) {
-                const_offset += nir_intrinsic_base(instr);
-
-                /* Shared variables have no buffer index, and all start from a
-                 * common base that we set up at the start of dispatch
+        } else if (is_shared_or_scratch) {
+                /* Shared and scratch variables have no buffer index, and all
+                 * start from a common base that we set up at the start of
+                 * dispatch.
                  */
-                offset = c->cs_shared_offset;
+                if (instr->intrinsic == nir_intrinsic_load_scratch ||
+                    instr->intrinsic == nir_intrinsic_store_scratch) {
+                        offset = c->spill_base;
+                } else {
+                        offset = c->cs_shared_offset;
+                        const_offset += nir_intrinsic_base(instr);
+                }
         } else {
                 offset = vir_uniform(c, QUNIFORM_SSBO_OFFSET,
                                      nir_src_as_uint(instr->src[is_store ?
@@ -1629,6 +1638,8 @@ ntq_emit_intrinsic(struct v3d_compile *c, nir_intrinsic_instr *instr)
         case nir_intrinsic_shared_atomic_comp_swap:
         case nir_intrinsic_load_shared:
         case nir_intrinsic_store_shared:
+        case nir_intrinsic_load_scratch:
+        case nir_intrinsic_store_scratch:
                 ntq_emit_tmu_general(c, instr, true);
                 break;
 
@@ -2308,6 +2319,11 @@ nir_to_vir(struct v3d_compile *c)
                 break;
         }
 
+        if (c->s->scratch_size) {
+                v3d_setup_spill_base(c);
+                c->spill_size += V3D_CHANNELS * c->s->scratch_size;
+        }
+
         if (c->s->info.stage == MESA_SHADER_FRAGMENT)
                 ntq_setup_fs_inputs(c);
         else
@@ -2524,7 +2540,7 @@ v3d_nir_to_vir(struct v3d_compile *c)
                         vir_remove_thrsw(c);
         }
 
-        if (c->spill_size &&
+        if (c->spills &&
             (V3D_DEBUG & (V3D_DEBUG_VIR |
                           v3d_debug_flag_for_shader_stage(c->s->info.stage)))) {
                 fprintf(stderr, "%s prog %d/%d spilled VIR:\n",
index b2bc40b10fe3df3ca089f96d563171d907e28934..2bc07de7e1f3ed31ee52e1b149e0ebbbaef43423 100644 (file)
@@ -749,6 +749,7 @@ struct qreg vir_uniform(struct v3d_compile *c,
                         enum quniform_contents contents,
                         uint32_t data);
 void vir_schedule_instructions(struct v3d_compile *c);
+void v3d_setup_spill_base(struct v3d_compile *c);
 struct v3d_qpu_instr v3d_qpu_nop(void);
 
 struct qreg vir_emit_def(struct v3d_compile *c, struct qinst *inst);
@@ -796,6 +797,7 @@ bool vir_opt_small_immediates(struct v3d_compile *c);
 bool vir_opt_vpm(struct v3d_compile *c);
 void v3d_nir_lower_blend(nir_shader *s, struct v3d_compile *c);
 void v3d_nir_lower_io(nir_shader *s, struct v3d_compile *c);
+void v3d_nir_lower_scratch(nir_shader *s);
 void v3d_nir_lower_txf_ms(nir_shader *s, struct v3d_compile *c);
 void v3d_nir_lower_image_load_store(nir_shader *s);
 void vir_lower_uniforms(struct v3d_compile *c);
diff --git a/src/broadcom/compiler/v3d_nir_lower_scratch.c b/src/broadcom/compiler/v3d_nir_lower_scratch.c
new file mode 100644 (file)
index 0000000..d23b8be
--- /dev/null
@@ -0,0 +1,153 @@
+/*
+ * Copyright © 2018 Intel Corporation
+ * Copyright © 2018 Broadcom
+ *
+ * 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 "v3d_compiler.h"
+#include "compiler/nir/nir_builder.h"
+#include "compiler/nir/nir_format_convert.h"
+
+/** @file v3d_nir_lower_scratch.c
+ *
+ * Swizzles around the addresses of
+ * nir_intrinsic_load_scratch/nir_intrinsic_store_scratch so that a QPU stores
+ * a cacheline at a time per dword of scratch access, scalarizing and removing
+ * writemasks in the process.
+ */
+
+static nir_ssa_def *
+v3d_nir_scratch_offset(nir_builder *b, nir_intrinsic_instr *instr)
+{
+        bool is_store = instr->intrinsic == nir_intrinsic_store_scratch;
+        nir_ssa_def *offset = nir_ssa_for_src(b, instr->src[is_store ? 1 : 0], 1);
+
+        assert(nir_intrinsic_align_mul(instr) >= 4);
+        assert(nir_intrinsic_align_offset(instr) == 0);
+
+        /* The spill_offset register will already have the subgroup ID (EIDX)
+         * shifted and ORed in at bit 2, so all we need to do is to move the
+         * dword index up above V3D_CHANNELS.
+         */
+        return nir_imul_imm(b, offset, V3D_CHANNELS);
+}
+
+static void
+v3d_nir_lower_load_scratch(nir_builder *b, nir_intrinsic_instr *instr)
+{
+        b->cursor = nir_before_instr(&instr->instr);
+
+        nir_ssa_def *offset = v3d_nir_scratch_offset(b,instr);
+
+        nir_ssa_def *chans[NIR_MAX_VEC_COMPONENTS];
+        for (int i = 0; i < instr->num_components; i++) {
+                nir_ssa_def *chan_offset =
+                        nir_iadd_imm(b, offset, V3D_CHANNELS * i * 4);
+
+                nir_intrinsic_instr *chan_instr =
+                        nir_intrinsic_instr_create(b->shader, instr->intrinsic);
+                chan_instr->num_components = 1;
+                nir_ssa_dest_init(&chan_instr->instr, &chan_instr->dest, 1,
+                                  instr->dest.ssa.bit_size, NULL);
+
+                chan_instr->src[0] = nir_src_for_ssa(chan_offset);
+
+                nir_intrinsic_set_align(chan_instr, 4, 0);
+
+                nir_builder_instr_insert(b, &chan_instr->instr);
+
+                chans[i] = &chan_instr->dest.ssa;
+        }
+
+        nir_ssa_def *result = nir_vec(b, chans, instr->num_components);
+        nir_ssa_def_rewrite_uses(&instr->dest.ssa, nir_src_for_ssa(result));
+        nir_instr_remove(&instr->instr);
+}
+
+static void
+v3d_nir_lower_store_scratch(nir_builder *b, nir_intrinsic_instr *instr)
+{
+        b->cursor = nir_before_instr(&instr->instr);
+
+        nir_ssa_def *offset = v3d_nir_scratch_offset(b, instr);
+        nir_ssa_def *value = nir_ssa_for_src(b, instr->src[0],
+                                             instr->num_components);
+
+        for (int i = 0; i < instr->num_components; i++) {
+                if (!(nir_intrinsic_write_mask(instr) & (1 << i)))
+                        continue;
+
+                nir_ssa_def *chan_offset =
+                        nir_iadd_imm(b, offset, V3D_CHANNELS * i * 4);
+
+                nir_intrinsic_instr *chan_instr =
+                        nir_intrinsic_instr_create(b->shader, instr->intrinsic);
+                chan_instr->num_components = 1;
+
+                chan_instr->src[0] = nir_src_for_ssa(nir_channel(b,
+                                                                 value,
+                                                                 i));
+                chan_instr->src[1] = nir_src_for_ssa(chan_offset);
+                nir_intrinsic_set_write_mask(chan_instr, 0x1);
+                nir_intrinsic_set_align(chan_instr, 4, 0);
+
+                nir_builder_instr_insert(b, &chan_instr->instr);
+        }
+
+        nir_instr_remove(&instr->instr);
+}
+
+void
+v3d_nir_lower_scratch(nir_shader *s)
+{
+        nir_foreach_function(function, s) {
+                if (!function->impl)
+                        continue;
+
+                nir_builder b;
+                nir_builder_init(&b, function->impl);
+
+                nir_foreach_block(block, function->impl) {
+                        nir_foreach_instr_safe(instr, block) {
+                                if (instr->type != nir_instr_type_intrinsic)
+                                        continue;
+
+                                nir_intrinsic_instr *intr =
+                                        nir_instr_as_intrinsic(instr);
+
+                                switch (intr->intrinsic) {
+                                case nir_intrinsic_load_scratch:
+                                        v3d_nir_lower_load_scratch(&b, intr);
+                                        break;
+                                case nir_intrinsic_store_scratch:
+                                        v3d_nir_lower_store_scratch(&b, intr);
+                                        break;
+                                default:
+                                        break;
+                                }
+                        }
+                }
+
+                nir_metadata_preserve(function->impl,
+                                      nir_metadata_block_index |
+                                      nir_metadata_dominance);
+        }
+}
index 6655e5e73bcd11e40a7152b541a84a8d73d39984..5d4e5dd103ada140e009fd6d70c01aceb4bea40d 100644 (file)
@@ -579,6 +579,12 @@ v3d_lower_nir(struct v3d_compile *c)
 
         NIR_PASS_V(c->s, nir_lower_tex, &tex_options);
         NIR_PASS_V(c->s, nir_lower_system_values);
+
+        NIR_PASS_V(c->s, nir_lower_vars_to_scratch,
+                   nir_var_function_temp,
+                   0,
+                   glsl_get_natural_size_align_bytes);
+        NIR_PASS_V(c->s, v3d_nir_lower_scratch);
 }
 
 static void
index 1c79ac93530d0a0c40ad67e27a43a835ebff4b80..32b9a8abe8e6d1849e415bdea48645cca0840009 100644 (file)
@@ -156,7 +156,7 @@ v3d_choose_spill_node(struct v3d_compile *c, struct ra_graph *g,
 /* The spill offset for this thread takes a bit of setup, so do it once at
  * program start.
  */
-static void
+void
 v3d_setup_spill_base(struct v3d_compile *c)
 {
         c->cursor = vir_before_block(vir_entry_block(c));
@@ -185,6 +185,8 @@ v3d_setup_spill_base(struct v3d_compile *c)
         /* Make sure that we don't spill the spilling setup instructions. */
         for (int i = start_num_temps; i < c->num_temps; i++)
                 BITSET_CLEAR(c->spillable, i);
+
+        c->cursor = vir_after_block(c->cur_block);
 }
 
 static void
index 6f91e35521a16f72ae895c0abbc85e61e71e7a02..0d9184279dff11d9f984474a228c39118430e7c4 100644 (file)
@@ -310,8 +310,9 @@ v3d_screen_get_shader_param(struct pipe_screen *pscreen, unsigned shader,
                 return 0;
         case PIPE_SHADER_CAP_INDIRECT_INPUT_ADDR:
         case PIPE_SHADER_CAP_INDIRECT_OUTPUT_ADDR:
-        case PIPE_SHADER_CAP_INDIRECT_TEMP_ADDR:
                 return 0;
+        case PIPE_SHADER_CAP_INDIRECT_TEMP_ADDR:
+                return 1;
         case PIPE_SHADER_CAP_INDIRECT_CONST_ADDR:
                 return 1;
         case PIPE_SHADER_CAP_SUBROUTINES: