[Mesa-dev] [PATCH 071/133] nir: Add a fused multiply-add peephole
Jason Ekstrand
jason at jlekstrand.net
Wed Dec 17 17:35:33 PST 2014
On Wed, Dec 17, 2014 at 5:13 PM, Connor Abbott <cwabbott0 at gmail.com> wrote:
>
> Would it be possible to drop this patch since this all gets deleted later?
>
We could, but I'd rather leave it in the history.
--Jason
>
> On Tue, Dec 16, 2014 at 1:05 AM, Jason Ekstrand <jason at jlekstrand.net>
> wrote:
> > ---
> > src/glsl/Makefile.sources | 1 +
> > src/glsl/nir/nir.h | 1 +
> > src/glsl/nir/nir_opcodes.h | 1 +
> > src/glsl/nir/nir_opt_peephole_ffma.c | 191
> +++++++++++++++++++++++++++++++
> > src/mesa/drivers/dri/i965/brw_fs_nir.cpp | 2 +
> > 5 files changed, 196 insertions(+)
> > create mode 100644 src/glsl/nir/nir_opt_peephole_ffma.c
> >
> > diff --git a/src/glsl/Makefile.sources b/src/glsl/Makefile.sources
> > index 38ffb85..31608dc 100644
> > --- a/src/glsl/Makefile.sources
> > +++ b/src/glsl/Makefile.sources
> > @@ -32,6 +32,7 @@ NIR_FILES = \
> > $(GLSL_SRCDIR)/nir/nir_opt_copy_propagate.c \
> > $(GLSL_SRCDIR)/nir/nir_opt_dce.c \
> > $(GLSL_SRCDIR)/nir/nir_opt_global_to_local.c \
> > + $(GLSL_SRCDIR)/nir/nir_opt_peephole_ffma.c \
> > $(GLSL_SRCDIR)/nir/nir_opt_peephole_select.c \
> > $(GLSL_SRCDIR)/nir/nir_print.c \
> > $(GLSL_SRCDIR)/nir/nir_remove_dead_variables.c \
> > diff --git a/src/glsl/nir/nir.h b/src/glsl/nir/nir.h
> > index bc2280b..e20eb7c 100644
> > --- a/src/glsl/nir/nir.h
> > +++ b/src/glsl/nir/nir.h
> > @@ -1374,6 +1374,7 @@ bool nir_opt_dce_impl(nir_function_impl *impl);
> > bool nir_opt_dce(nir_shader *shader);
> >
> > bool nir_opt_peephole_select(nir_shader *shader);
> > +bool nir_opt_peephole_ffma(nir_shader *shader);
> >
> > #ifdef __cplusplus
> > } /* extern "C" */
> > diff --git a/src/glsl/nir/nir_opcodes.h b/src/glsl/nir/nir_opcodes.h
> > index 988f691..35d4634 100644
> > --- a/src/glsl/nir/nir_opcodes.h
> > +++ b/src/glsl/nir/nir_opcodes.h
> > @@ -307,6 +307,7 @@ BINOP_HORIZ(vec2, 2, nir_type_unsigned, 1,
> nir_type_unsigned, 1, nir_type_unsign
> > ARR(src1_size, src2_size, src3_size), \
> > ARR(nir_type_unsigned, nir_type_unsigned, nir_type_unsigned))
> >
> > +/* fma(a, b, c) = (a * b) + c */
> > TRIOP(ffma, nir_type_float)
> >
> > TRIOP(flrp, nir_type_float)
> > diff --git a/src/glsl/nir/nir_opt_peephole_ffma.c
> b/src/glsl/nir/nir_opt_peephole_ffma.c
> > new file mode 100644
> > index 0000000..2c9b8e5
> > --- /dev/null
> > +++ b/src/glsl/nir/nir_opt_peephole_ffma.c
> > @@ -0,0 +1,191 @@
> > +/*
> > + * Copyright © 2014 Intel Corporation
> > + *
> > + * 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:
> > + * Jason Ekstrand (jason at jlekstrand.net)
> > + *
> > + */
> > +
> > +#include "nir.h"
> > +
> > +/*
> > + * Implements a small peephole optimization that looks for a multiply
> that
> > + * is only ever used in an add and replaces both with an fma.
> > + */
> > +
> > +struct peephole_ffma_state {
> > + void *mem_ctx;
> > + nir_function_impl *impl;
> > + bool progress;
> > +};
> > +
> > +static inline nir_alu_instr *
> > +get_mul_for_src(nir_alu_instr *add, unsigned idx)
> > +{
> > + if (!add->src[idx].src.is_ssa)
> > + return NULL;
> > +
> > + /* We can't handle these in between the operations */
> > + if (add->src[idx].negate || add->src[idx].abs)
> > + return NULL;
> > +
> > + nir_instr *instr = add->src[idx].src.ssa->parent_instr;
> > + if (instr->type != nir_instr_type_alu)
> > + return NULL;
> > +
> > + nir_alu_instr *mul = nir_instr_as_alu(instr);
> > + if (mul->op != nir_op_fmul)
> > + return NULL;
> > +
> > + /* Can't handle a saturate in between */
> > + if (mul->dest.saturate)
> > + return NULL;
> > +
> > + /* We already know that the same source is not used twice in the add
> and
> > + * we will assume valid use-def information, so this check is
> sufficient
> > + */
> > + if (mul->dest.dest.ssa.uses->entries > 1)
> > + return NULL; /* Not the only use */
> > +
> > + return mul;
> > +}
> > +
> > +/* Copies (and maybe swizzles) the given ALU source */
> > +static inline void
> > +copy_alu_src(void *mem_ctx, nir_alu_src *new_src, nir_alu_src old_src,
> > + uint8_t *swizzle)
> > +{
> > + new_src->src = nir_src_copy(old_src.src, mem_ctx);
> > + new_src->abs = old_src.abs;
> > + new_src->negate = old_src.negate;
> > +
> > + if (swizzle == NULL) {
> > + memcpy(new_src->swizzle, old_src.swizzle, sizeof old_src.swizzle);
> > + } else {
> > + for (int i = 0; i < 4; ++i) {
> > + if (swizzle[i] < 4)
> > + new_src->swizzle[i] = old_src.swizzle[swizzle[i]];
> > + }
> > + }
> > +}
> > +
> > +static bool
> > +nir_opt_peephole_ffma_block(nir_block *block, void *void_state)
> > +{
> > + struct peephole_ffma_state *state = void_state;
> > +
> > + nir_foreach_instr_safe(block, instr) {
> > + if (instr->type != nir_instr_type_alu)
> > + continue;
> > +
> > + nir_alu_instr *add = nir_instr_as_alu(instr);
> > + if (add->op != nir_op_fadd)
> > + continue;
> > +
> > + /* TODO: Maybe bail if this expression is considered "precise"? */
> > +
> > + /* This, is the case a + a. We would rather handle this with an
> > + * algebraic reduction than fuse it. Also, we want to only fuse
> > + * things where the multiply is used only once and, in this case,
> > + * it would be used twice by the same instruction.
> > + */
> > + if (add->src[0].src.is_ssa && add->src[1].src.is_ssa &&
> > + add->src[0].src.ssa == add->src[1].src.ssa)
> > + continue;
> > +
> > + nir_alu_instr *mul = get_mul_for_src(add, 0);
> > + unsigned mul_src = 0;
> > +
> > + if (mul == NULL) {
> > + mul = get_mul_for_src(add, 1);
> > + mul_src = 1;
> > + }
> > +
> > + if (mul == NULL)
> > + continue;
> > +
> > + nir_alu_instr *ffma = nir_alu_instr_create(state->mem_ctx,
> nir_op_ffma);
> > + ffma->dest.saturate = add->dest.saturate;
> > + ffma->dest.write_mask = add->dest.write_mask;
> > +
> > + copy_alu_src(state->mem_ctx, &ffma->src[0], mul->src[0],
> > + add->src[mul_src].swizzle);
> > + copy_alu_src(state->mem_ctx, &ffma->src[1], mul->src[1],
> > + add->src[mul_src].swizzle);
> > + copy_alu_src(state->mem_ctx, &ffma->src[2], add->src[1 -
> mul_src], NULL);
> > +
> > + if (add->dest.dest.is_ssa) {
> > + ffma->dest.dest.is_ssa = true;
> > + nir_ssa_def_init(state->impl, &ffma->instr,
> &ffma->dest.dest.ssa,
> > + add->dest.dest.ssa.num_components,
> > + add->dest.dest.ssa.name);
> > +
> > + nir_src ffma_dest_src = {
> > + .is_ssa = true,
> > + .ssa = &ffma->dest.dest.ssa,
> > + };
> > + nir_ssa_def_rewrite_uses(&add->dest.dest.ssa, ffma_dest_src,
> > + state->mem_ctx);
> > + } else {
> > + ffma->dest.dest = nir_dest_copy(add->dest.dest,
> state->mem_ctx);
> > + }
> > +
> > + nir_instr_insert_before(&add->instr, &ffma->instr);
> > + nir_instr_remove(&add->instr);
> > + nir_instr_remove(&mul->instr);
> > +
> > + state->progress = true;
> > + }
> > +
> > + return true;
> > +}
> > +
> > +static bool
> > +nir_opt_peephole_ffma_impl(nir_function_impl *impl)
> > +{
> > + struct peephole_ffma_state state;
> > +
> > + state.mem_ctx = ralloc_parent(impl);
> > + state.impl = impl;
> > + state.progress = false;
> > +
> > + nir_foreach_block(impl, nir_opt_peephole_ffma_block, &state);
> > +
> > + if (state.progress)
> > + nir_metadata_dirty(impl, nir_metadata_block_index |
> > + nir_metadata_dominance);
> > +
> > + return state.progress;
> > +}
> > +
> > +bool
> > +nir_opt_peephole_ffma(nir_shader *shader)
> > +{
> > + bool progress = false;
> > +
> > + nir_foreach_overload(shader, overload) {
> > + if (overload->impl)
> > + progress |= nir_opt_peephole_ffma_impl(overload->impl);
> > + }
> > +
> > + return progress;
> > +}
> > diff --git a/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
> b/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
> > index b2c5863..b6720af 100644
> > --- a/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
> > +++ b/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
> > @@ -63,6 +63,8 @@ fs_visitor::emit_nir_code()
> > nir_validate_shader(nir);
> > nir_opt_peephole_select(nir);
> > nir_validate_shader(nir);
> > + nir_opt_peephole_ffma(nir);
> > + nir_validate_shader(nir);
> > nir_print_shader(nir, stderr);
> >
> > nir_convert_from_ssa(nir);
> > --
> > 2.2.0
> >
> > _______________________________________________
> > mesa-dev mailing list
> > mesa-dev at lists.freedesktop.org
> > http://lists.freedesktop.org/mailman/listinfo/mesa-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.freedesktop.org/archives/mesa-dev/attachments/20141217/cd932df3/attachment.html>
More information about the mesa-dev
mailing list