[Mesa-dev] [PATCH 1/7] mesa: implement ARB/KHR_parallel_shader_compile
Ian Romanick
idr at freedesktop.org
Thu Jan 3 19:40:24 UTC 2019
On 11/28/18 6:59 PM, Marek Olšák wrote:
> From: Marek Olšák <marek.olsak at amd.com>
>
> Tested by piglit.
It doesn't look like there are any piglit test
> ---
> docs/features.txt | 2 +-
> docs/relnotes/19.0.0.html | 2 ++
> src/mapi/glapi/gen/gl_API.xml | 15 ++++++++++++++-
> src/mesa/main/dd.h | 7 +++++++
> src/mesa/main/extensions_table.h | 2 ++
> src/mesa/main/get_hash_params.py | 3 +++
> src/mesa/main/hint.c | 12 ++++++++++++
> src/mesa/main/hint.h | 4 ++++
> src/mesa/main/mtypes.h | 1 +
> src/mesa/main/shaderapi.c | 10 ++++++++++
> src/mesa/main/tests/dispatch_sanity.cpp | 4 ++++
> 11 files changed, 60 insertions(+), 2 deletions(-)
>
> diff --git a/docs/features.txt b/docs/features.txt
> index 8999e42519c..7b827de6a92 100644
> --- a/docs/features.txt
> +++ b/docs/features.txt
> @@ -295,21 +295,21 @@ GLES3.2, GLSL ES 3.2 -- all DONE: i965/gen9+, radeonsi, virgl
> GL_OES_texture_storage_multisample_2d_array DONE (all drivers that support GL_ARB_texture_multisample)
>
> Khronos, ARB, and OES extensions that are not part of any OpenGL or OpenGL ES version:
>
> GL_ARB_bindless_texture DONE (nvc0, radeonsi)
> GL_ARB_cl_event not started
> GL_ARB_compute_variable_group_size DONE (nvc0, radeonsi)
> GL_ARB_ES3_2_compatibility DONE (i965/gen8+, radeonsi, virgl)
> GL_ARB_fragment_shader_interlock DONE (i965)
> GL_ARB_gpu_shader_int64 DONE (i965/gen8+, nvc0, radeonsi, softpipe, llvmpipe)
> - GL_ARB_parallel_shader_compile not started, but Chia-I Wu did some related work in 2014
> + GL_ARB_parallel_shader_compile DONE (all drivers)
> GL_ARB_post_depth_coverage DONE (i965, nvc0)
> GL_ARB_robustness_isolation not started
> GL_ARB_sample_locations DONE (nvc0)
> GL_ARB_seamless_cubemap_per_texture DONE (freedreno, i965, nvc0, radeonsi, r600, softpipe, swr, virgl)
> GL_ARB_shader_ballot DONE (i965/gen8+, nvc0, radeonsi)
> GL_ARB_shader_clock DONE (i965/gen7+, nv50, nvc0, r600, radeonsi, virgl)
> GL_ARB_shader_stencil_export DONE (i965/gen9+, r600, radeonsi, softpipe, llvmpipe, swr, virgl)
> GL_ARB_shader_viewport_layer_array DONE (i965/gen6+, nvc0, radeonsi)
> GL_ARB_sparse_buffer DONE (radeonsi/CIK+)
> GL_ARB_sparse_texture not started
> diff --git a/docs/relnotes/19.0.0.html b/docs/relnotes/19.0.0.html
> index bc1776e8f4e..540482bca5f 100644
> --- a/docs/relnotes/19.0.0.html
> +++ b/docs/relnotes/19.0.0.html
> @@ -33,24 +33,26 @@ Compatibility contexts may report a lower version depending on each driver.
> <h2>SHA256 checksums</h2>
> <pre>
> TBD.
> </pre>
>
>
> <h2>New features</h2>
>
> <ul>
> <li>GL_AMD_texture_texture4 on all GL 4.0 drivers.</li>
> +<li>GL_ARB_parallel_shader_compile on all drivers.</li>
> <li>GL_EXT_shader_implicit_conversions on all drivers (ES extension).</li>
> <li>GL_EXT_texture_compression_bptc on all GL 4.0 drivers (ES extension).<li>
> <li>GL_EXT_texture_compression_rgtc on all GL 3.0 drivers (ES extension).<li>
> <li>GL_EXT_texture_view on drivers supporting texture views (ES extension).<li>
> +<li>GL_KHR_parallel_shader_compile on all drivers.</li>
> <li>GL_OES_texture_view on drivers supporting texture views (ES extension).</li>
> </ul>
>
> <h2>Bug fixes</h2>
>
> <ul>
> <li>TBD</li>
> </ul>
>
> <h2>Changes</h2>
> diff --git a/src/mapi/glapi/gen/gl_API.xml b/src/mapi/glapi/gen/gl_API.xml
> index f4d0808f13b..4ce691b361b 100644
> --- a/src/mapi/glapi/gen/gl_API.xml
> +++ b/src/mapi/glapi/gen/gl_API.xml
> @@ -8402,21 +8402,34 @@
> <param name="maxX" type="GLfloat"/>
> <param name="maxY" type="GLfloat"/>
> <param name="maxZ" type="GLfloat"/>
> <param name="maxW" type="GLfloat"/>
> </function>
>
> </category>
>
> <xi:include href="ARB_gpu_shader_int64.xml" xmlns:xi="http://www.w3.org/2001/XInclude"/>
>
> -<!-- ARB extension 179 - 189 -->
> +<category name="ARB_parallel_shader_compile" number="179">
> + <enum name="MAX_SHADER_COMPILER_THREADS_ARB" value="0x91B0"/>
> + <enum name="COMPLETION_STATUS_ARB" value="0x91B1"/>
> +
> + <function name="MaxShaderCompilerThreadsKHR" es2="2.0">
> + <param name="count" type="GLuint"/>
> + </function>
> +
> + <function name="MaxShaderCompilerThreadsARB" alias="MaxShaderCompilerThreadsKHR">
> + <param name="count" type="GLuint"/>
> + </function>
> +</category>
> +
> +<!-- ARB extension 180 - 189 -->
>
> <xi:include href="ARB_gl_spirv.xml" xmlns:xi="http://www.w3.org/2001/XInclude"/>
>
> <!-- Non-ARB extensions sorted by extension number. -->
>
> <category name="GL_EXT_blend_color" number="2">
> <enum name="CONSTANT_COLOR_EXT" value="0x8001"/>
> <enum name="ONE_MINUS_CONSTANT_COLOR_EXT" value="0x8002"/>
> <enum name="CONSTANT_ALPHA_EXT" value="0x8003"/>
> <enum name="ONE_MINUS_CONSTANT_ALPHA_EXT" value="0x8004"/>
> diff --git a/src/mesa/main/dd.h b/src/mesa/main/dd.h
> index f14c3e04e91..92b6ecac33c 100644
> --- a/src/mesa/main/dd.h
> +++ b/src/mesa/main/dd.h
> @@ -1292,20 +1292,27 @@ struct dd_function_table {
> /**
> * Called to initialize gl_program::driver_cache_blob (and size) with a
> * ralloc allocated buffer.
> *
> * This buffer will be saved and restored as part of the gl_program
> * serialization and deserialization.
> */
> void (*ShaderCacheSerializeDriverBlob)(struct gl_context *ctx,
> struct gl_program *prog);
> /*@}*/
> +
> + /**
> + * \name Set the number of compiler threads for ARB_parallel_shader_compile
> + */
> + void (*SetMaxShaderCompilerThreads)(struct gl_context *ctx, unsigned count);
> + bool (*GetShaderProgramCompletionStatus)(struct gl_context *ctx,
> + struct gl_shader_program *shprog);
> };
>
>
> /**
> * Per-vertex functions.
> *
> * These are the functions which can appear between glBegin and glEnd.
> * Depending on whether we're inside or outside a glBegin/End pair
> * and whether we're in immediate mode or building a display list, these
> * functions behave differently. This structure allows us to switch
> diff --git a/src/mesa/main/extensions_table.h b/src/mesa/main/extensions_table.h
> index 2a3e1ab32ca..075d72f16c2 100644
> --- a/src/mesa/main/extensions_table.h
> +++ b/src/mesa/main/extensions_table.h
> @@ -92,20 +92,21 @@ EXT(ARB_internalformat_query , ARB_internalformat_query
> EXT(ARB_internalformat_query2 , ARB_internalformat_query2 , GLL, GLC, x , x , 2013)
> EXT(ARB_invalidate_subdata , dummy_true , GLL, GLC, x , x , 2012)
> EXT(ARB_map_buffer_alignment , dummy_true , GLL, GLC, x , x , 2011)
> EXT(ARB_map_buffer_range , ARB_map_buffer_range , GLL, GLC, x , x , 2008)
> EXT(ARB_multi_bind , dummy_true , GLL, GLC, x , x , 2013)
> EXT(ARB_multi_draw_indirect , ARB_draw_indirect , GLL, GLC, x , x , 2012)
> EXT(ARB_multisample , dummy_true , GLL, x , x , x , 1994)
> EXT(ARB_multitexture , dummy_true , GLL, x , x , x , 1998)
> EXT(ARB_occlusion_query , ARB_occlusion_query , GLL, x , x , x , 2001)
> EXT(ARB_occlusion_query2 , ARB_occlusion_query2 , GLL, GLC, x , x , 2003)
> +EXT(ARB_parallel_shader_compile , dummy_true , GLL, GLC, x , x , 2015)
> EXT(ARB_pipeline_statistics_query , ARB_pipeline_statistics_query , GLL, GLC, x , x , 2014)
> EXT(ARB_pixel_buffer_object , EXT_pixel_buffer_object , GLL, GLC, x , x , 2004)
> EXT(ARB_point_parameters , EXT_point_parameters , GLL, x , x , x , 1997)
> EXT(ARB_point_sprite , ARB_point_sprite , GLL, GLC, x , x , 2003)
> EXT(ARB_polygon_offset_clamp , ARB_polygon_offset_clamp , GLL, GLC, x , x , 2017)
> EXT(ARB_post_depth_coverage , ARB_post_depth_coverage , GLL, GLC, x , x, 2015)
> EXT(ARB_program_interface_query , dummy_true , GLL, GLC, x , x , 2012)
> EXT(ARB_provoking_vertex , EXT_provoking_vertex , GLL, GLC, x , x , 2009)
> EXT(ARB_query_buffer_object , ARB_query_buffer_object , GLL, GLC, x , x , 2013)
> EXT(ARB_robust_buffer_access_behavior , ARB_robust_buffer_access_behavior , GLL, GLC, x , x , 2012)
> @@ -324,20 +325,21 @@ EXT(INGR_blend_func_separate , EXT_blend_func_separate
> EXT(INTEL_conservative_rasterization , INTEL_conservative_rasterization , x , GLC, x , 31, 2013)
> EXT(INTEL_fragment_shader_ordering , INTEL_fragment_shader_ordering , GLL, GLC, x , x , 2013)
> EXT(INTEL_performance_query , INTEL_performance_query , GLL, GLC, x , ES2, 2013)
> EXT(INTEL_shader_atomic_float_minmax , INTEL_shader_atomic_float_minmax , GLL, GLC, x , x , 2018)
>
> EXT(KHR_blend_equation_advanced , KHR_blend_equation_advanced , GLL, GLC, x , ES2, 2014)
> EXT(KHR_blend_equation_advanced_coherent , KHR_blend_equation_advanced_coherent , GLL, GLC, x , ES2, 2014)
> EXT(KHR_context_flush_control , dummy_true , GLL, GLC, x , ES2, 2014)
> EXT(KHR_debug , dummy_true , GLL, GLC, 11, ES2, 2012)
> EXT(KHR_no_error , dummy_true , GLL, GLC, ES1, ES2, 2015)
> +EXT(KHR_parallel_shader_compile , dummy_true , GLL, GLC, x , ES2, 2017)
> EXT(KHR_robust_buffer_access_behavior , ARB_robust_buffer_access_behavior , GLL, GLC, x , ES2, 2014)
> EXT(KHR_robustness , KHR_robustness , GLL, GLC, x , ES2, 2012)
> EXT(KHR_texture_compression_astc_hdr , KHR_texture_compression_astc_hdr , GLL, GLC, x , ES2, 2012)
> EXT(KHR_texture_compression_astc_ldr , KHR_texture_compression_astc_ldr , GLL, GLC, x , ES2, 2012)
> EXT(KHR_texture_compression_astc_sliced_3d , KHR_texture_compression_astc_sliced_3d , GLL, GLC, x , ES2, 2015)
>
> EXT(MESA_framebuffer_flip_y , MESA_framebuffer_flip_y , x, x, x , 31, 2018)
> EXT(MESA_pack_invert , MESA_pack_invert , GLL, GLC, x , x , 2002)
> EXT(MESA_shader_integer_functions , MESA_shader_integer_functions , GLL, GLC, x , 30, 2016)
> EXT(MESA_texture_signed_rgba , EXT_texture_snorm , GLL, GLC, x , x , 2009)
> diff --git a/src/mesa/main/get_hash_params.py b/src/mesa/main/get_hash_params.py
> index 0b8ec9ea0bb..579c6e459aa 100644
> --- a/src/mesa/main/get_hash_params.py
> +++ b/src/mesa/main/get_hash_params.py
> @@ -273,20 +273,23 @@ descriptor=[
> [ "PACK_SKIP_ROWS", "CONTEXT_INT(Pack.SkipRows), NO_EXTRA" ],
> [ "UNPACK_ROW_LENGTH", "CONTEXT_INT(Unpack.RowLength), NO_EXTRA" ],
> [ "UNPACK_SKIP_PIXELS", "CONTEXT_INT(Unpack.SkipPixels), NO_EXTRA" ],
> [ "UNPACK_SKIP_ROWS", "CONTEXT_INT(Unpack.SkipRows), NO_EXTRA" ],
> [ "UNPACK_SKIP_IMAGES", "CONTEXT_INT(Unpack.SkipImages), NO_EXTRA" ],
> [ "UNPACK_IMAGE_HEIGHT", "CONTEXT_INT(Unpack.ImageHeight), NO_EXTRA" ],
>
> # GL_ARB_draw_buffers
> [ "MAX_DRAW_BUFFERS_ARB", "CONTEXT_INT(Const.MaxDrawBuffers), NO_EXTRA" ],
>
> +# GL_ARB_parallel_shader_compile
> + [ "MAX_SHADER_COMPILER_THREADS_ARB", "CONTEXT_INT(Hint.MaxShaderCompilerThreads), NO_EXTRA" ],
> +
> # GL_EXT_framebuffer_object / GL_NV_fbo_color_attachments
> [ "MAX_COLOR_ATTACHMENTS", "CONTEXT_INT(Const.MaxColorAttachments), NO_EXTRA" ],
>
> # GL_ARB_draw_buffers / GL_NV_draw_buffers (for ES 2.0)
> [ "DRAW_BUFFER0_ARB", "BUFFER_ENUM16(ColorDrawBuffer[0]), NO_EXTRA" ],
> [ "DRAW_BUFFER1_ARB", "BUFFER_ENUM16(ColorDrawBuffer[1]), extra_valid_draw_buffer" ],
> [ "DRAW_BUFFER2_ARB", "BUFFER_ENUM16(ColorDrawBuffer[2]), extra_valid_draw_buffer" ],
> [ "DRAW_BUFFER3_ARB", "BUFFER_ENUM16(ColorDrawBuffer[3]), extra_valid_draw_buffer" ],
> [ "DRAW_BUFFER4_ARB", "BUFFER_ENUM16(ColorDrawBuffer[4]), extra_valid_draw_buffer" ],
> [ "DRAW_BUFFER5_ARB", "BUFFER_ENUM16(ColorDrawBuffer[5]), extra_valid_draw_buffer" ],
> diff --git a/src/mesa/main/hint.c b/src/mesa/main/hint.c
> index 5d0c15d35ab..2bcaeebc4aa 100644
> --- a/src/mesa/main/hint.c
> +++ b/src/mesa/main/hint.c
> @@ -123,27 +123,39 @@ _mesa_Hint( GLenum target, GLenum mode )
> default:
> goto invalid_target;
> }
> return;
>
> invalid_target:
> _mesa_error(ctx, GL_INVALID_ENUM, "glHint(target)");
> return;
> }
>
> +/* GL_ARB_parallel_shader_compile */
> +void GLAPIENTRY
> +_mesa_MaxShaderCompilerThreadsKHR(GLuint count)
> +{
> + GET_CURRENT_CONTEXT(ctx);
> +
> + ctx->Hint.MaxShaderCompilerThreads = count;
> +
> + if (ctx->Driver.SetMaxShaderCompilerThreads)
> + ctx->Driver.SetMaxShaderCompilerThreads(ctx, count);
> +}
>
> /**********************************************************************/
> /***** Initialization *****/
> /**********************************************************************/
>
> void _mesa_init_hint( struct gl_context * ctx )
> {
> /* Hint group */
> ctx->Hint.PerspectiveCorrection = GL_DONT_CARE;
> ctx->Hint.PointSmooth = GL_DONT_CARE;
> ctx->Hint.LineSmooth = GL_DONT_CARE;
> ctx->Hint.PolygonSmooth = GL_DONT_CARE;
> ctx->Hint.Fog = GL_DONT_CARE;
> ctx->Hint.TextureCompression = GL_DONT_CARE;
> ctx->Hint.GenerateMipmap = GL_DONT_CARE;
> ctx->Hint.FragmentShaderDerivative = GL_DONT_CARE;
> + ctx->Hint.MaxShaderCompilerThreads = 0xffffffff;
> }
> diff --git a/src/mesa/main/hint.h b/src/mesa/main/hint.h
> index 87febfeecc0..2d7c710af26 100644
> --- a/src/mesa/main/hint.h
> +++ b/src/mesa/main/hint.h
> @@ -36,14 +36,18 @@
> #ifndef HINT_H
> #define HINT_H
>
> #include "glheader.h"
>
> struct gl_context;
>
> extern void GLAPIENTRY
> _mesa_Hint( GLenum target, GLenum mode );
>
> +/* GL_KHR_parallel_shader_compile */
> +extern void GLAPIENTRY
> +_mesa_MaxShaderCompilerThreadsKHR(GLuint count);
> +
> extern void
> _mesa_init_hint( struct gl_context * ctx );
>
> #endif
> diff --git a/src/mesa/main/mtypes.h b/src/mesa/main/mtypes.h
> index 157d45bc0ba..5b2b29d1f5d 100644
> --- a/src/mesa/main/mtypes.h
> +++ b/src/mesa/main/mtypes.h
> @@ -596,20 +596,21 @@ struct gl_fog_attrib
> struct gl_hint_attrib
> {
> GLenum16 PerspectiveCorrection;
> GLenum16 PointSmooth;
> GLenum16 LineSmooth;
> GLenum16 PolygonSmooth;
> GLenum16 Fog;
> GLenum16 TextureCompression; /**< GL_ARB_texture_compression */
> GLenum16 GenerateMipmap; /**< GL_SGIS_generate_mipmap */
> GLenum16 FragmentShaderDerivative; /**< GL_ARB_fragment_shader */
> + GLuint MaxShaderCompilerThreads; /**< GL_ARB_parallel_shader_compile */
> };
>
>
> /**
> * Lighting attribute group (GL_LIGHT_BIT).
> */
> struct gl_light_attrib
> {
> struct gl_light Light[MAX_LIGHTS]; /**< Array of light sources */
> struct gl_lightmodel Model; /**< Lighting model */
> diff --git a/src/mesa/main/shaderapi.c b/src/mesa/main/shaderapi.c
> index 2ea8d965aba..01342c04e8f 100644
> --- a/src/mesa/main/shaderapi.c
> +++ b/src/mesa/main/shaderapi.c
> @@ -682,20 +682,26 @@ get_programiv(struct gl_context *ctx, GLuint program, GLenum pname,
> || _mesa_is_gles3(ctx);
>
> if (!shProg) {
> return;
> }
>
> switch (pname) {
> case GL_DELETE_STATUS:
> *params = shProg->DeletePending;
> return;
> + case GL_COMPLETION_STATUS_ARB:
> + if (ctx->Driver.GetShaderProgramCompletionStatus)
> + *params = ctx->Driver.GetShaderProgramCompletionStatus(ctx, shProg);
> + else
> + *params = GL_TRUE;
> + return;
> case GL_LINK_STATUS:
> *params = shProg->data->LinkStatus ? GL_TRUE : GL_FALSE;
> return;
> case GL_VALIDATE_STATUS:
> *params = shProg->data->Validated;
> return;
> case GL_INFO_LOG_LENGTH:
> *params = (shProg->data->InfoLog && shProg->data->InfoLog[0] != '\0') ?
> strlen(shProg->data->InfoLog) + 1 : 0;
> return;
> @@ -953,20 +959,24 @@ get_shaderiv(struct gl_context *ctx, GLuint name, GLenum pname, GLint *params)
> return;
> }
>
> switch (pname) {
> case GL_SHADER_TYPE:
> *params = shader->Type;
> break;
> case GL_DELETE_STATUS:
> *params = shader->DeletePending;
> break;
> + case GL_COMPLETION_STATUS_ARB:
> + /* _mesa_glsl_compile_shader is not offloaded to other threads. */
> + *params = GL_TRUE;
> + return;
So... glComplieShader is completely serial, but glLinkShader can operate
in parallel? That differs quite a bit from what Chia-I was doing... and
explains why a lot the threading fixes he had is absent here. Most of
the expensive work happens in glLinkShader, so this seems reasonable.
Have you been able to measure this on the start-up time of any real apps?
> case GL_COMPILE_STATUS:
> *params = shader->CompileStatus ? GL_TRUE : GL_FALSE;
> break;
> case GL_INFO_LOG_LENGTH:
> *params = (shader->InfoLog && shader->InfoLog[0] != '\0') ?
> strlen(shader->InfoLog) + 1 : 0;
> break;
> case GL_SHADER_SOURCE_LENGTH:
> *params = shader->Source ? strlen((char *) shader->Source) + 1 : 0;
> break;
> diff --git a/src/mesa/main/tests/dispatch_sanity.cpp b/src/mesa/main/tests/dispatch_sanity.cpp
> index fb2acfbdeea..6aa6b87882e 100644
> --- a/src/mesa/main/tests/dispatch_sanity.cpp
> +++ b/src/mesa/main/tests/dispatch_sanity.cpp
> @@ -1191,20 +1191,22 @@ const struct function common_desktop_functions_possible[] = {
> { "glEvaluateDepthValuesARB", 30, -1 },
>
> /* GL_ARB_indirect_parameters */
> { "glMultiDrawArraysIndirectCountARB", 11, -1 },
> { "glMultiDrawElementsIndirectCountARB", 11, -1 },
>
> /* GL_AMD_framebuffer_multisample_advanced */
> { "glRenderbufferStorageMultisampleAdvancedAMD", 11, -1 },
> { "glNamedRenderbufferStorageMultisampleAdvancedAMD", 11, -1 },
>
> + { "glMaxShaderCompilerThreadsKHR", 11, -1 },
> +
> { NULL, 0, -1 }
> };
>
> const struct function gl_compatibility_functions_possible[] = {
> { "glNewList", 10, _gloffset_NewList },
> { "glEndList", 10, _gloffset_EndList },
> { "glCallList", 10, _gloffset_CallList },
> { "glCallLists", 10, _gloffset_CallLists },
> { "glDeleteLists", 10, _gloffset_DeleteLists },
> { "glGenLists", 10, _gloffset_GenLists },
> @@ -2229,20 +2231,22 @@ const struct function gles2_functions_possible[] = {
>
> /* GL_NV_conservative_raster */
> { "glSubpixelPrecisionBiasNV", 20, -1 },
>
> /* GL_NV_conservative_raster_dilate */
> { "glConservativeRasterParameterfNV", 20, -1 },
>
> /* GL_NV_conservative_raster_pre_snap_triangles */
> { "glConservativeRasterParameteriNV", 20, -1 },
>
> + { "glMaxShaderCompilerThreadsKHR", 20, -1 },
> +
> { NULL, 0, -1 }
> };
>
> const struct function gles3_functions_possible[] = {
> // We check for the aliased -EXT version in GLES 2
> // { "glBeginQuery", 30, -1 },
> { "glBeginTransformFeedback", 30, -1 },
> { "glBindBufferBase", 30, -1 },
> { "glBindBufferRange", 30, -1 },
> { "glBindSampler", 30, -1 },
>
More information about the mesa-dev
mailing list