st/glsl: start spilling out common st glsl conversion code
authorTimothy Arceri <tarceri@itsqueeze.com>
Wed, 4 Apr 2018 06:01:21 +0000 (16:01 +1000)
committerTimothy Arceri <tarceri@itsqueeze.com>
Wed, 6 Mar 2019 23:05:20 +0000 (23:05 +0000)
The NIR and TGSI paths are currently intertwined which makes it
not only hard to follow but also makes it hard to take advantage
of the differences in IR.

Here we take the first step to splitting that path apart. With
this we take the opportunity to no longer call the GLSL IR
optimisation passes after the final lowering calls for NIR. We
can instead just use the NIR passes which can produce better code
and should also result in faster compile times.

The speed-up can be measured in some dolphin uber shaders due to
no longer calling lower_if_to_cond_assign() for example
dolphin/ubershaders/120.shader_test goes from ~1.63 -> ~1.53
seconds on my machine.

There are some code changes as a result of not calling
lower_if_to_cond_assign(), this is because it flattens ifs that
contain UBOs where as NIR's peephole select doesn't. This is
were most of the regressions in Max Waves happens with shader-db.

shader-db results (VEGA):

Totals from affected shaders:
SGPRS: 2349056 -> 2349640 (0.02 %)
VGPRS: 1322160 -> 1323300 (0.09 %)
Spilled SGPRs: 21190 -> 21527 (1.59 %)
Spilled VGPRs: 99 -> 99 (0.00 %)
Private memory VGPRs: 0 -> 0 (0.00 %)
Scratch size: 72 -> 72 (0.00 %) dwords per thread
Code Size: 57260904 -> 57270932 (0.02 %) bytes
Compile Time: 1107186 -> 1022942 (-7.61 %) milliseconds
LDS: 786 -> 786 (0.00 %) blocks
Max Waves: 391932 -> 391619 (-0.08 %)
Wait states: 0 -> 0 (0.00 %)

Reviewed-by: Eric Anholt <eric@anholt.net>
src/mesa/Makefile.sources
src/mesa/meson.build
src/mesa/state_tracker/st_cb_program.c
src/mesa/state_tracker/st_glsl_to_ir.cpp [new file with mode: 0644]
src/mesa/state_tracker/st_glsl_to_ir.h [new file with mode: 0644]
src/mesa/state_tracker/st_glsl_to_tgsi.cpp
src/mesa/state_tracker/st_glsl_to_tgsi.h

index f482963a1acfc3b1e8cb35ac198effe956200054..d554408d01b3519f55cb320572565ef1f6068871 100644 (file)
@@ -521,6 +521,8 @@ STATETRACKER_FILES = \
        state_tracker/st_gen_mipmap.c \
        state_tracker/st_gen_mipmap.h \
        state_tracker/st_gl_api.h \
+       state_tracker/st_glsl_to_ir.cpp \
+       state_tracker/st_glsl_to_ir.h \
        state_tracker/st_glsl_to_nir.cpp \
        state_tracker/st_glsl_to_tgsi.cpp \
        state_tracker/st_glsl_to_tgsi.h \
index 517506b12c4ace062664e345815d64b1bad37981..05a8bb5e10033b751c6c1d46fc2b9f8cb115a172 100644 (file)
@@ -565,6 +565,8 @@ files_libmesa_gallium = files(
   'state_tracker/st_gen_mipmap.c',
   'state_tracker/st_gen_mipmap.h',
   'state_tracker/st_gl_api.h',
+  'state_tracker/st_glsl_to_ir.cpp',
+  'state_tracker/st_glsl_to_ir.h',
   'state_tracker/st_glsl_to_nir.cpp',
   'state_tracker/st_glsl_to_tgsi.cpp',
   'state_tracker/st_glsl_to_tgsi.h',
index 555fc5d5ad9bc3c7c4f55a5d7b91d4c7448d0a8d..825a09d7fdab2410832a10db9d088ba99227c7eb 100644 (file)
@@ -45,7 +45,7 @@
 #include "st_program.h"
 #include "st_mesa_to_tgsi.h"
 #include "st_cb_program.h"
-#include "st_glsl_to_tgsi.h"
+#include "st_glsl_to_ir.h"
 #include "st_atifs_to_tgsi.h"
 
 
diff --git a/src/mesa/state_tracker/st_glsl_to_ir.cpp b/src/mesa/state_tracker/st_glsl_to_ir.cpp
new file mode 100644 (file)
index 0000000..ea56bf4
--- /dev/null
@@ -0,0 +1,172 @@
+/*
+ * Copyright (C) 2005-2007  Brian Paul   All Rights Reserved.
+ * Copyright (C) 2008  VMware, Inc.   All Rights Reserved.
+ * Copyright © 2010 Intel Corporation
+ * Copyright © 2011 Bryan Cain
+ *
+ * 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 "compiler/glsl/glsl_parser_extras.h"
+#include "compiler/glsl/ir_optimization.h"
+#include "compiler/glsl/program.h"
+
+#include "st_nir.h"
+#include "st_shader_cache.h"
+#include "st_glsl_to_tgsi.h"
+
+#include "tgsi/tgsi_from_mesa.h"
+
+extern "C" {
+
+/**
+ * Link a shader.
+ * Called via ctx->Driver.LinkShader()
+ * This is a shared function that branches off to either GLSL IR -> TGSI or
+ * GLSL IR -> NIR
+ */
+GLboolean
+st_link_shader(struct gl_context *ctx, struct gl_shader_program *prog)
+{
+   struct pipe_screen *pscreen = ctx->st->pipe->screen;
+
+   enum pipe_shader_ir preferred_ir = (enum pipe_shader_ir)
+      pscreen->get_shader_param(pscreen, PIPE_SHADER_VERTEX,
+                                PIPE_SHADER_CAP_PREFERRED_IR);
+   bool use_nir = preferred_ir == PIPE_SHADER_IR_NIR;
+
+   /* Return early if we are loading the shader from on-disk cache */
+   if (st_load_ir_from_disk_cache(ctx, prog, use_nir)) {
+      return GL_TRUE;
+   }
+
+   assert(prog->data->LinkStatus);
+
+   for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
+      if (prog->_LinkedShaders[i] == NULL)
+         continue;
+
+      struct gl_linked_shader *shader = prog->_LinkedShaders[i];
+      exec_list *ir = shader->ir;
+      gl_shader_stage stage = shader->Stage;
+      const struct gl_shader_compiler_options *options =
+            &ctx->Const.ShaderCompilerOptions[stage];
+
+      /* If there are forms of indirect addressing that the driver
+       * cannot handle, perform the lowering pass.
+       */
+      if (options->EmitNoIndirectInput || options->EmitNoIndirectOutput ||
+          options->EmitNoIndirectTemp || options->EmitNoIndirectUniform) {
+         lower_variable_index_to_cond_assign(stage, ir,
+                                             options->EmitNoIndirectInput,
+                                             options->EmitNoIndirectOutput,
+                                             options->EmitNoIndirectTemp,
+                                             options->EmitNoIndirectUniform);
+      }
+
+      enum pipe_shader_type ptarget = pipe_shader_type_from_mesa(stage);
+      bool have_dround = pscreen->get_shader_param(pscreen, ptarget,
+                                                   PIPE_SHADER_CAP_TGSI_DROUND_SUPPORTED);
+      bool have_dfrexp = pscreen->get_shader_param(pscreen, ptarget,
+                                                   PIPE_SHADER_CAP_TGSI_DFRACEXP_DLDEXP_SUPPORTED);
+      bool have_ldexp = pscreen->get_shader_param(pscreen, ptarget,
+                                                  PIPE_SHADER_CAP_TGSI_LDEXP_SUPPORTED);
+
+      if (!pscreen->get_param(pscreen, PIPE_CAP_INT64_DIVMOD))
+         lower_64bit_integer_instructions(ir, DIV64 | MOD64);
+
+      if (ctx->Extensions.ARB_shading_language_packing) {
+         unsigned lower_inst = LOWER_PACK_SNORM_2x16 |
+                               LOWER_UNPACK_SNORM_2x16 |
+                               LOWER_PACK_UNORM_2x16 |
+                               LOWER_UNPACK_UNORM_2x16 |
+                               LOWER_PACK_SNORM_4x8 |
+                               LOWER_UNPACK_SNORM_4x8 |
+                               LOWER_UNPACK_UNORM_4x8 |
+                               LOWER_PACK_UNORM_4x8;
+
+         if (ctx->Extensions.ARB_gpu_shader5)
+            lower_inst |= LOWER_PACK_USE_BFI |
+                          LOWER_PACK_USE_BFE;
+         if (!ctx->st->has_half_float_packing)
+            lower_inst |= LOWER_PACK_HALF_2x16 |
+                          LOWER_UNPACK_HALF_2x16;
+
+         lower_packing_builtins(ir, lower_inst);
+      }
+
+      if (!pscreen->get_param(pscreen, PIPE_CAP_TEXTURE_GATHER_OFFSETS))
+         lower_offset_arrays(ir);
+      do_mat_op_to_vec(ir);
+
+      if (stage == MESA_SHADER_FRAGMENT)
+         lower_blend_equation_advanced(
+            shader, ctx->Extensions.KHR_blend_equation_advanced_coherent);
+
+      lower_instructions(ir,
+                         MOD_TO_FLOOR |
+                         FDIV_TO_MUL_RCP |
+                         EXP_TO_EXP2 |
+                         LOG_TO_LOG2 |
+                         MUL64_TO_MUL_AND_MUL_HIGH |
+                         (have_ldexp ? 0 : LDEXP_TO_ARITH) |
+                         (have_dfrexp ? 0 : DFREXP_DLDEXP_TO_ARITH) |
+                         CARRY_TO_ARITH |
+                         BORROW_TO_ARITH |
+                         (have_dround ? 0 : DOPS_TO_DFRAC) |
+                         (options->EmitNoPow ? POW_TO_EXP2 : 0) |
+                         (!ctx->Const.NativeIntegers ? INT_DIV_TO_MUL_RCP : 0) |
+                         (options->EmitNoSat ? SAT_TO_CLAMP : 0) |
+                         (ctx->Const.ForceGLSLAbsSqrt ? SQRT_TO_ABS_SQRT : 0) |
+                         /* Assume that if ARB_gpu_shader5 is not supported
+                          * then all of the extended integer functions need
+                          * lowering.  It may be necessary to add some caps
+                          * for individual instructions.
+                          */
+                         (!ctx->Extensions.ARB_gpu_shader5
+                          ? BIT_COUNT_TO_MATH |
+                            EXTRACT_TO_SHIFTS |
+                            INSERT_TO_SHIFTS |
+                            REVERSE_TO_SHIFTS |
+                            FIND_LSB_TO_FLOAT_CAST |
+                            FIND_MSB_TO_FLOAT_CAST |
+                            IMUL_HIGH_TO_MUL
+                          : 0));
+
+      do_vec_index_to_cond_assign(ir);
+      lower_vector_insert(ir, true);
+      lower_quadop_vector(ir, false);
+      lower_noise(ir);
+      if (options->MaxIfDepth == 0) {
+         lower_discard(ir);
+      }
+
+      validate_ir_tree(ir);
+   }
+
+   build_program_resource_list(ctx, prog);
+
+   if (use_nir)
+      return st_link_nir(ctx, prog);
+   else
+      return st_link_tgsi(ctx, prog);
+}
+
+} /* extern "C" */
diff --git a/src/mesa/state_tracker/st_glsl_to_ir.h b/src/mesa/state_tracker/st_glsl_to_ir.h
new file mode 100644 (file)
index 0000000..1239c66
--- /dev/null
@@ -0,0 +1,36 @@
+/*
+ * Copyright © 2010 Intel Corporation
+ * Copyright © 2011 Bryan Cain
+ *
+ * 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 "main/mtypes.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+GLboolean
+st_link_shader(struct gl_context *ctx, struct gl_shader_program *prog);
+
+#ifdef __cplusplus
+}
+#endif
index 5b7203542f39fa916d3b5af831dafb8c21f04d8e..f8771125f4569ce88644ece23aa20d38cf7e12d6 100644 (file)
@@ -53,8 +53,6 @@
 #include "st_program.h"
 #include "st_mesa_to_tgsi.h"
 #include "st_format.h"
-#include "st_nir.h"
-#include "st_shader_cache.h"
 #include "st_glsl_to_tgsi_temprename.h"
 
 #include "util/hash_table.h"
@@ -7287,131 +7285,29 @@ has_unsupported_control_flow(exec_list *ir,
    return visitor.unsupported;
 }
 
-extern "C" {
-
 /**
  * Link a shader.
- * Called via ctx->Driver.LinkShader()
  * This actually involves converting GLSL IR into an intermediate TGSI-like IR
  * with code lowering and other optimizations.
  */
 GLboolean
-st_link_shader(struct gl_context *ctx, struct gl_shader_program *prog)
+st_link_tgsi(struct gl_context *ctx, struct gl_shader_program *prog)
 {
    struct pipe_screen *pscreen = ctx->st->pipe->screen;
 
-   enum pipe_shader_ir preferred_ir = (enum pipe_shader_ir)
-      pscreen->get_shader_param(pscreen, PIPE_SHADER_VERTEX,
-                                PIPE_SHADER_CAP_PREFERRED_IR);
-   bool use_nir = preferred_ir == PIPE_SHADER_IR_NIR;
-
-   /* Return early if we are loading the shader from on-disk cache */
-   if (st_load_ir_from_disk_cache(ctx, prog, use_nir)) {
-      return GL_TRUE;
-   }
-
-   assert(prog->data->LinkStatus);
-
    for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
-      if (prog->_LinkedShaders[i] == NULL)
+      struct gl_linked_shader *shader = prog->_LinkedShaders[i];
+      if (shader == NULL)
          continue;
 
-      struct gl_linked_shader *shader = prog->_LinkedShaders[i];
       exec_list *ir = shader->ir;
       gl_shader_stage stage = shader->Stage;
+      enum pipe_shader_type ptarget = pipe_shader_type_from_mesa(stage);
       const struct gl_shader_compiler_options *options =
             &ctx->Const.ShaderCompilerOptions[stage];
-      enum pipe_shader_type ptarget = pipe_shader_type_from_mesa(stage);
-      bool have_dround = pscreen->get_shader_param(pscreen, ptarget,
-                                                   PIPE_SHADER_CAP_TGSI_DROUND_SUPPORTED);
-      bool have_dfrexp = pscreen->get_shader_param(pscreen, ptarget,
-                                                   PIPE_SHADER_CAP_TGSI_DFRACEXP_DLDEXP_SUPPORTED);
-      bool have_ldexp = pscreen->get_shader_param(pscreen, ptarget,
-                                                  PIPE_SHADER_CAP_TGSI_LDEXP_SUPPORTED);
+
       unsigned if_threshold = pscreen->get_shader_param(pscreen, ptarget,
                                                         PIPE_SHADER_CAP_LOWER_IF_THRESHOLD);
-
-      /* If there are forms of indirect addressing that the driver
-       * cannot handle, perform the lowering pass.
-       */
-      if (options->EmitNoIndirectInput || options->EmitNoIndirectOutput ||
-          options->EmitNoIndirectTemp || options->EmitNoIndirectUniform) {
-         lower_variable_index_to_cond_assign(stage, ir,
-                                             options->EmitNoIndirectInput,
-                                             options->EmitNoIndirectOutput,
-                                             options->EmitNoIndirectTemp,
-                                             options->EmitNoIndirectUniform);
-      }
-
-      if (!pscreen->get_param(pscreen, PIPE_CAP_INT64_DIVMOD))
-         lower_64bit_integer_instructions(ir, DIV64 | MOD64);
-
-      if (ctx->Extensions.ARB_shading_language_packing) {
-         unsigned lower_inst = LOWER_PACK_SNORM_2x16 |
-                               LOWER_UNPACK_SNORM_2x16 |
-                               LOWER_PACK_UNORM_2x16 |
-                               LOWER_UNPACK_UNORM_2x16 |
-                               LOWER_PACK_SNORM_4x8 |
-                               LOWER_UNPACK_SNORM_4x8 |
-                               LOWER_UNPACK_UNORM_4x8 |
-                               LOWER_PACK_UNORM_4x8;
-
-         if (ctx->Extensions.ARB_gpu_shader5)
-            lower_inst |= LOWER_PACK_USE_BFI |
-                          LOWER_PACK_USE_BFE;
-         if (!ctx->st->has_half_float_packing)
-            lower_inst |= LOWER_PACK_HALF_2x16 |
-                          LOWER_UNPACK_HALF_2x16;
-
-         lower_packing_builtins(ir, lower_inst);
-      }
-
-      if (!pscreen->get_param(pscreen, PIPE_CAP_TEXTURE_GATHER_OFFSETS))
-         lower_offset_arrays(ir);
-      do_mat_op_to_vec(ir);
-
-      if (stage == MESA_SHADER_FRAGMENT)
-         lower_blend_equation_advanced(
-            shader, ctx->Extensions.KHR_blend_equation_advanced_coherent);
-
-      lower_instructions(ir,
-                         MOD_TO_FLOOR |
-                         FDIV_TO_MUL_RCP |
-                         EXP_TO_EXP2 |
-                         LOG_TO_LOG2 |
-                         MUL64_TO_MUL_AND_MUL_HIGH |
-                         (have_ldexp ? 0 : LDEXP_TO_ARITH) |
-                         (have_dfrexp ? 0 : DFREXP_DLDEXP_TO_ARITH) |
-                         CARRY_TO_ARITH |
-                         BORROW_TO_ARITH |
-                         (have_dround ? 0 : DOPS_TO_DFRAC) |
-                         (options->EmitNoPow ? POW_TO_EXP2 : 0) |
-                         (!ctx->Const.NativeIntegers ? INT_DIV_TO_MUL_RCP : 0) |
-                         (options->EmitNoSat ? SAT_TO_CLAMP : 0) |
-                         (ctx->Const.ForceGLSLAbsSqrt ? SQRT_TO_ABS_SQRT : 0) |
-                         /* Assume that if ARB_gpu_shader5 is not supported
-                          * then all of the extended integer functions need
-                          * lowering.  It may be necessary to add some caps
-                          * for individual instructions.
-                          */
-                         (!ctx->Extensions.ARB_gpu_shader5
-                          ? BIT_COUNT_TO_MATH |
-                            EXTRACT_TO_SHIFTS |
-                            INSERT_TO_SHIFTS |
-                            REVERSE_TO_SHIFTS |
-                            FIND_LSB_TO_FLOAT_CAST |
-                            FIND_MSB_TO_FLOAT_CAST |
-                            IMUL_HIGH_TO_MUL
-                          : 0));
-
-      do_vec_index_to_cond_assign(ir);
-      lower_vector_insert(ir, true);
-      lower_quadop_vector(ir, false);
-      lower_noise(ir);
-      if (options->MaxIfDepth == 0) {
-         lower_discard(ir);
-      }
-
       if (ctx->Const.GLSLOptimizeConservatively) {
          /* Do it once and repeat only if there's unsupported control flow. */
          do {
@@ -7437,17 +7333,6 @@ st_link_shader(struct gl_context *ctx, struct gl_shader_program *prog)
       do_vec_index_to_cond_assign(ir);
 
       validate_ir_tree(ir);
-   }
-
-   build_program_resource_list(ctx, prog);
-
-   if (use_nir)
-      return st_link_nir(ctx, prog);
-
-   for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
-      struct gl_linked_shader *shader = prog->_LinkedShaders[i];
-      if (shader == NULL)
-         continue;
 
       struct gl_program *linked_prog =
          get_mesa_program_tgsi(ctx, prog, shader);
@@ -7466,6 +7351,8 @@ st_link_shader(struct gl_context *ctx, struct gl_shader_program *prog)
    return GL_TRUE;
 }
 
+extern "C" {
+
 void
 st_translate_stream_output_info(struct gl_transform_feedback_info *info,
                                 const ubyte outputMapping[],
index c8a87abca41fbf8ec1f8e33887bc8015ac17f4df..fb89c39fb8873ebe143b0cfd87b684c108d3d9f7 100644 (file)
@@ -58,7 +58,8 @@ enum pipe_error st_translate_program(
 
 void free_glsl_to_tgsi_visitor(struct glsl_to_tgsi_visitor *v);
 
-GLboolean st_link_shader(struct gl_context *ctx, struct gl_shader_program *prog);
+GLboolean
+st_link_tgsi(struct gl_context *ctx, struct gl_shader_program *prog);
 
 void
 st_translate_stream_output_info(struct gl_transform_feedback_info *info,