[Mesa-dev] [PATCH 2/2] i965: Add ARB_fragment_shader_interlock support.
Francisco Jerez
currojerez at riseup.net
Wed Apr 19 23:18:02 UTC 2017
Hi Pam, looks good overall, a couple of comments below,
Plamena Manolova <plamena.manolova at intel.com> writes:
> Adds suppport for ARB_fragment_shader_interlock. We achieve
> the interlock and fragment ordering by issuing a memory fence
> via sendc.
>
> Signed-off-by: Plamena Manolova <plamena.manolova at intel.com>
> ---
> docs/features.txt | 2 +-
> docs/relnotes/17.1.0.html | 1 +
> src/intel/compiler/brw_eu.h | 4 +++
> src/intel/compiler/brw_eu_defines.h | 2 ++
> src/intel/compiler/brw_eu_emit.c | 47 ++++++++++++++++++++++++++++
> src/intel/compiler/brw_fs_generator.cpp | 4 +++
> src/intel/compiler/brw_fs_nir.cpp | 15 +++++++++
> src/intel/compiler/brw_shader.cpp | 4 +++
> src/mesa/drivers/dri/i965/intel_extensions.c | 5 +++
> 9 files changed, 83 insertions(+), 1 deletion(-)
>
> diff --git a/docs/features.txt b/docs/features.txt
> index 5f63632..a6237c0 100644
> --- a/docs/features.txt
> +++ b/docs/features.txt
> @@ -281,7 +281,7 @@ Khronos, ARB, and OES extensions that are not part of any OpenGL or OpenGL ES ve
> GL_ARB_cl_event not started
> GL_ARB_compute_variable_group_size DONE (nvc0, radeonsi)
> GL_ARB_ES3_2_compatibility DONE (i965/gen8+)
> - GL_ARB_fragment_shader_interlock not started
> + GL_ARB_fragment_shader_interlock DONE (i965)
> GL_ARB_gl_spirv not started
> GL_ARB_gpu_shader_int64 DONE (i965/gen8+, nvc0, radeonsi, softpipe, llvmpipe)
> GL_ARB_indirect_parameters DONE (nvc0, radeonsi)
> diff --git a/docs/relnotes/17.1.0.html b/docs/relnotes/17.1.0.html
> index e7cfe38..1b2393f 100644
> --- a/docs/relnotes/17.1.0.html
> +++ b/docs/relnotes/17.1.0.html
> @@ -45,6 +45,7 @@ Note: some of the new features are only available with certain drivers.
>
> <ul>
> <li>OpenGL 4.2 on i965/ivb</li>
> +<li>GL_ARB_fragment_shader_interlock on i965</li>
> <li>GL_ARB_gpu_shader_fp64 on i965/ivybridge</li>
> <li>GL_ARB_gpu_shader_int64 on i965/gen8+, nvc0, radeonsi, softpipe, llvmpipe</li>
> <li>GL_ARB_shader_ballot on nvc0, radeonsi</li>
> diff --git a/src/intel/compiler/brw_eu.h b/src/intel/compiler/brw_eu.h
> index f422595..117cfae 100644
> --- a/src/intel/compiler/brw_eu.h
> +++ b/src/intel/compiler/brw_eu.h
> @@ -480,6 +480,10 @@ brw_memory_fence(struct brw_codegen *p,
> struct brw_reg dst);
>
> void
> +brw_interlock(struct brw_codegen *p,
> + struct brw_reg dst);
> +
> +void
> brw_pixel_interpolator_query(struct brw_codegen *p,
> struct brw_reg dest,
> struct brw_reg mrf,
> diff --git a/src/intel/compiler/brw_eu_defines.h b/src/intel/compiler/brw_eu_defines.h
> index 13a70f6..9eb5210 100644
> --- a/src/intel/compiler/brw_eu_defines.h
> +++ b/src/intel/compiler/brw_eu_defines.h
> @@ -444,6 +444,8 @@ enum opcode {
> */
> SHADER_OPCODE_BROADCAST,
>
> + SHADER_OPCODE_INTERLOCK,
> +
> VEC4_OPCODE_MOV_BYTES,
> VEC4_OPCODE_PACK_BYTES,
> VEC4_OPCODE_UNPACK_UNIFORM,
> diff --git a/src/intel/compiler/brw_eu_emit.c b/src/intel/compiler/brw_eu_emit.c
> index 231d6fd..52adf22 100644
> --- a/src/intel/compiler/brw_eu_emit.c
> +++ b/src/intel/compiler/brw_eu_emit.c
> @@ -3403,6 +3403,53 @@ brw_memory_fence(struct brw_codegen *p,
> }
>
> void
> +brw_interlock(struct brw_codegen *p,
> + struct brw_reg dst)
> +{
> + const struct gen_device_info *devinfo = p->devinfo;
> + const bool commit_enable = devinfo->gen == 7 && !devinfo->is_haswell;
> + struct brw_inst *insn;
> +
> + brw_push_insn_state(p);
> + brw_set_default_mask_control(p, BRW_MASK_DISABLE);
> + brw_set_default_exec_size(p, BRW_EXECUTE_1);
> + dst = vec1(dst);
> +
> + /* Set dst as destination for dependency tracking, the MEMORY_FENCE
> + * message doesn't write anything back.
> + */
> + /* BRW_OPCODE_SENDC is what the interlock actually depends on */
> + insn = next_insn(p, BRW_OPCODE_SENDC);
> + dst = retype(dst, BRW_REGISTER_TYPE_UW);
> + brw_set_dest(p, insn, dst);
> + brw_set_src0(p, insn, dst);
> + /* Issuing a memory fence ensures the ordering of fragments */
> + brw_set_memory_fence_message(p, insn, GEN7_SFID_DATAPORT_DATA_CACHE,
> + commit_enable);
> +
> + if (devinfo->gen == 7 && !devinfo->is_haswell) {
> + /* IVB does typed surface access through the render cache, so we need to
> + * flush it too. Use a different register so both flushes can be
> + * pipelined by the hardware.
> + */
> + insn = next_insn(p, BRW_OPCODE_SENDC);
> + brw_set_dest(p, insn, offset(dst, 1));
> + brw_set_src0(p, insn, offset(dst, 1));
> + brw_set_memory_fence_message(p, insn, GEN6_SFID_DATAPORT_RENDER_CACHE,
> + commit_enable);
> +
> + /* Now write the response of the second message into the response of the
> + * first to trigger a pipeline stall -- This way future render and data
> + * cache messages will be properly ordered with respect to past data and
> + * render cache messages.
> + */
> + brw_MOV(p, dst, offset(dst, 1));
> + }
> +
> + brw_pop_insn_state(p);
> +}
> +
AFICT this is basically a copy of brw_memory_fence() with s/SEND/SENDC/.
It may be a good idea to add an opcode argument to brw_memory_fence()
instead and drop brw_interlock() altogether so we don't have to worry
about keeping the two barrier implementations in sync if e.g. fixes are
applied in the future.
> +void
> brw_pixel_interpolator_query(struct brw_codegen *p,
> struct brw_reg dest,
> struct brw_reg mrf,
> diff --git a/src/intel/compiler/brw_fs_generator.cpp b/src/intel/compiler/brw_fs_generator.cpp
> index a7f95cc..093c12d 100644
> --- a/src/intel/compiler/brw_fs_generator.cpp
> +++ b/src/intel/compiler/brw_fs_generator.cpp
> @@ -2070,6 +2070,10 @@ fs_generator::generate_code(const cfg_t *cfg, int dispatch_width)
> brw_memory_fence(p, dst);
> break;
>
> + case SHADER_OPCODE_INTERLOCK:
> + brw_interlock(p, dst);
> + break;
> +
> case SHADER_OPCODE_FIND_LIVE_CHANNEL: {
> const struct brw_reg mask =
> brw_stage_has_packed_dispatch(devinfo, stage,
> diff --git a/src/intel/compiler/brw_fs_nir.cpp b/src/intel/compiler/brw_fs_nir.cpp
> index 23cd4b7..9d6302c 100644
> --- a/src/intel/compiler/brw_fs_nir.cpp
> +++ b/src/intel/compiler/brw_fs_nir.cpp
> @@ -4138,6 +4138,21 @@ fs_visitor::nir_emit_intrinsic(const fs_builder &bld, nir_intrinsic_instr *instr
> break;
> }
>
> + case nir_intrinsic_begin_invocation_interlock_ARB: {
> + const fs_builder ubld = bld.group(8, 0);
> + const fs_reg tmp = ubld.vgrf(BRW_REGISTER_TYPE_UD, 2);
> +
> + ubld.emit(SHADER_OPCODE_INTERLOCK, tmp)->size_written = 2 *
> + REG_SIZE;
> +
> + break;
> + }
> +
> + case nir_intrinsic_end_invocation_interlock_ARB: {
> + /* We don't need to do anything here */
> + break;
> + }
> +
> default:
> unreachable("unknown intrinsic");
> }
> diff --git a/src/intel/compiler/brw_shader.cpp b/src/intel/compiler/brw_shader.cpp
> index 304b4ec..8b9e42f 100644
> --- a/src/intel/compiler/brw_shader.cpp
> +++ b/src/intel/compiler/brw_shader.cpp
> @@ -290,6 +290,9 @@ brw_instruction_name(const struct gen_device_info *devinfo, enum opcode op)
> return "typed_surface_write_logical";
> case SHADER_OPCODE_MEMORY_FENCE:
> return "memory_fence";
> + case SHADER_OPCODE_INTERLOCK:
> + /* For an interlock we actually issue a memory fence via sendc. */
> + return "memory_fence";
Still because your "interlock" instruction has different semantics it
will be useful for debugging to print what IR instruction was actually
in the shader rather than lying to the user. :P
>
> case SHADER_OPCODE_LOAD_PAYLOAD:
> return "load_payload";
> @@ -989,6 +992,7 @@ backend_instruction::has_side_effects() const
> case SHADER_OPCODE_TYPED_SURFACE_WRITE:
> case SHADER_OPCODE_TYPED_SURFACE_WRITE_LOGICAL:
> case SHADER_OPCODE_MEMORY_FENCE:
> + case SHADER_OPCODE_INTERLOCK:
> case SHADER_OPCODE_URB_WRITE_SIMD8:
> case SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT:
> case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED:
> diff --git a/src/mesa/drivers/dri/i965/intel_extensions.c b/src/mesa/drivers/dri/i965/intel_extensions.c
> index 0133fa1..830e530 100644
> --- a/src/mesa/drivers/dri/i965/intel_extensions.c
> +++ b/src/mesa/drivers/dri/i965/intel_extensions.c
> @@ -297,4 +297,9 @@ intelInitExtensions(struct gl_context *ctx)
> ctx->Extensions.EXT_texture_compression_s3tc = true;
>
> ctx->Extensions.ANGLE_texture_compression_dxt = true;
> +
> + if ((ctx->Extensions.ARB_shader_image_load_store || ctx->Version >= 42) &&
> + ctx->Const.GLSLVersion >= 420) {
> + ctx->Extensions.ARB_fragment_shader_interlock = true;
Since this is basically enabling it if and only if GL 4.2 is supported
due to the GLSLVersion check, you could as well do a single
unconditional assignment "ctx->Extensions.ARB_fragment_shader_interlock
= true" around intel_extensions.c:232 in the same block as the
ARB_draw_indirect enable.
> + }
> }
> --
> 2.9.3
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 212 bytes
Desc: not available
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20170419/62057207/attachment.sig>
More information about the mesa-dev
mailing list