[Mesa-dev] [PATCH 3/3] gallium: Add tgsi_to_nir to get a nir_shader for a TGSI shader.
Kenneth Graunke
kenneth at whitecape.org
Fri Mar 27 15:02:25 PDT 2015
On Friday, March 27, 2015 01:54:32 PM Eric Anholt wrote:
> This will be used by the VC4 driver for doing device-independent
> optimization, and hopefully eventually replacing its whole IR. It also
> may be useful to other drivers for the same reason.
>
> v2: Add all of the instructions I was relying on tgsi_lowering to remove,
> and more.
> v3: Rebase on SSA rework of the builder.
> v4: Use the NIR ineg operation instead of doing a src modifier.
> v5: Don't use ineg for fnegs. (infer_src_type on MOV doesn't do what I
> expect, again).
> v6: Fix handling of multi-channel KILL_IF sources.
> v7: Make ttn_get_f() return a swizzle of a scalar load_const, rather than
> a vector load_const. CSE doesn't recognize that srcs out of those
> channels are actually all the same.
> v8: Rebase on nir_builder auto-sizing, make the scalar arguments to
> non-ALU instructions actually be scalars.
> v9: Add support for if/loop instructions, additional texture targets, and
> untested support for indirect addressing on temps.
> v10: Rebase on master, drop bad comment about control flow and just choose
> the X channel, use int comparison opcodes in LIT for now, drop unused
> pipe_context argument..
> v11: Fix translation of LRP (previously missed because I mis-translated
> back out), use nir_builder init helpers.
> v12: Rebase on master, adding explicit include of mtypes.h to get
> INTERP_QUALIFIER_*
> v13: Rebase on variables being in lists instead of hash tables, drop use
> of mtypes.h in favor of util/pipeline.h. Use Ken's nir_builder
> swizzle and fmov/imov_alu helpers, drop "struct" in front of
> nir_builder, use nir_builder directly as the function arg in a lot of
> cases, drop redundant members of ttn_compile that are also in
> nir_builder, drop some half-baked malloc failure handling.
> ---
>
> This series is present as part of my vc4-nir-rebase-qir-2 branch. The
> shader-db results across the branch are:
>
> total uniforms in shared programs: 13433 -> 13434 (0.01%)
> uniforms in affected programs: 62 -> 63 (1.61%)
> total instructions in shared programs: 40003 -> 39794 (-0.52%)
> instructions in affected programs: 15494 -> 15285 (-1.35%)
>
> I don't get to delete my driver's optimization code in that branch
> yet, notably because of optimization available in the programmable
> blending I have to generate.
Hi Eric!
I have a bunch of comments below, but overall this looks great.
You should probably have someone who knows TGSI better than I do review
it, but for what it's worth, this is:
Reviewed-by: Kenneth Graunke <kenneth at whitecape.org>
>
> src/gallium/auxiliary/Makefile.sources | 1 +
> src/gallium/auxiliary/nir/tgsi_to_nir.c | 1425
+++++++++++++++++++++++++++++++
> src/gallium/auxiliary/nir/tgsi_to_nir.h | 30 +
> 3 files changed, 1456 insertions(+)
> create mode 100644 src/gallium/auxiliary/nir/tgsi_to_nir.c
> create mode 100644 src/gallium/auxiliary/nir/tgsi_to_nir.h
>
> diff --git a/src/gallium/auxiliary/Makefile.sources
b/src/gallium/auxiliary/Makefile.sources
> index 09496fa..08e4e4c 100644
> --- a/src/gallium/auxiliary/Makefile.sources
> +++ b/src/gallium/auxiliary/Makefile.sources
> @@ -69,6 +69,7 @@ C_SOURCES := \
> indices/u_indices_priv.h \
> indices/u_primconvert.c \
> indices/u_primconvert.h \
> + nir/tgsi_to_nir.c \
> os/os_memory_aligned.h \
> os/os_memory_debug.h \
> os/os_memory_stdc.h \
> diff --git a/src/gallium/auxiliary/nir/tgsi_to_nir.c
b/src/gallium/auxiliary/nir/tgsi_to_nir.c
> new file mode 100644
> index 0000000..e8cf9f4
> --- /dev/null
> +++ b/src/gallium/auxiliary/nir/tgsi_to_nir.c
> @@ -0,0 +1,1425 @@
> +/*
> + * Copyright © 2014-2015 Broadcom
> + * Copyright (C) 2014 Rob Clark <robclark at freedesktop.org>
> + *
> + * 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 "util/ralloc.h"
> +#include "glsl/nir/nir.h"
> +#include "glsl/nir/nir_builder.h"
> +#include "glsl/list.h"
> +#include "util/pipeline.h"
> +
> +#include "nir/tgsi_to_nir.h"
> +#include "tgsi/tgsi_parse.h"
> +#include "tgsi/tgsi_dump.h"
> +#include "tgsi/tgsi_info.h"
> +#include "tgsi/tgsi_scan.h"
> +
> +#define SWIZ(X, Y, Z, W) (unsigned[4]){ \
> + TGSI_SWIZZLE_##X, \
> + TGSI_SWIZZLE_##Y, \
> + TGSI_SWIZZLE_##Z, \
> + TGSI_SWIZZLE_##W, \
> + }
> +
> +struct ttn_reg_info {
> + /** nir register containing this TGSI index. */
> + nir_register *reg;
> + /** Offset (in vec4s) from the start of var for this TGSI index. */
> + int offset;
> +};
> +
> +struct ttn_compile {
> + union tgsi_full_token *token;
> + nir_builder build;
> + struct nir_shader *s;
> + struct tgsi_shader_info *scan;
> +
> + struct ttn_reg_info *output_regs;
> + struct ttn_reg_info *temp_regs;
> + nir_ssa_def **imm_defs;
> +
> + nir_register *addr_reg;
> +
> + /**
> + * Stack of cf_node_lists where instructions should be pushed as we pop
> + * back out of the control flow stack.
> + *
> + * For each IF/ELSE/ENDIF block, if_stack[if_stack_pos] has where the
else
> + * instructions should be placed, and if_stack[if_stack_pos - 1] has
where
> + * the next instructions outside of the if/then/else block go.
> + */
> + struct exec_list **if_stack;
> + unsigned if_stack_pos;
> +
> + /**
> + * Stack of cf_node_lists where instructions should be pushed as we pop
> + * back out of the control flow stack.
> + *
> + * loop_stack[loop_stack_pos - 1] contains the cf_node_list for the
outside
> + * of the loop.
> + */
> + struct exec_list **loop_stack;
> + unsigned loop_stack_pos;
> +
> + /* How many TGSI_FILE_IMMEDIATE vec4s have been parsed so far. */
> + unsigned next_imm;
> +};
> +
> +#define ttn_swizzle(b, src, x, y, z, w) \
> + nir_swizzle(b, src, SWIZ(x, y, z, w), 4, false)
> +#define ttn_channel(b, src, swiz) \
> + nir_swizzle(b, src, SWIZ(swiz, swiz, swiz, swiz), 1, false)
> +
> +static nir_ssa_def *
> +ttn_src_for_dest(nir_builder *b, nir_alu_dest *dest)
> +{
> + nir_alu_src src;
> + memset(&src, 0, sizeof(src));
> +
> + if (dest->dest.is_ssa)
> + src.src = nir_src_for_ssa(&dest->dest.ssa);
> + else {
> + assert(!dest->dest.reg.indirect);
> + src.src = nir_src_for_reg(dest->dest.reg.reg);
> + src.src.reg.base_offset = dest->dest.reg.base_offset;
> + }
> +
> + for (int i = 0; i < 4; i++)
> + src.swizzle[i] = i;
> +
> + return nir_fmov_alu(b, src, 4);
> +}
> +
> +static void
> +ttn_emit_declaration(struct ttn_compile *c)
> +{
> + nir_builder *b = &c->build;
> + struct tgsi_full_declaration *decl = &c->token->FullDeclaration;
> + unsigned array_size = decl->Range.Last - decl->Range.First + 1;
> + unsigned file = decl->Declaration.File;
> + unsigned i;
> +
> + if (file == TGSI_FILE_TEMPORARY) {
> + nir_register *reg;
> + if (c->scan->indirect_files & (1 << file)) {
> + reg = nir_local_reg_create(b->impl);
> + reg->num_components = 4;
> + reg->num_array_elems = array_size;
> +
> + for (i = 0; i < array_size; i++) {
> + c->temp_regs[decl->Range.First + i].reg = reg;
> + c->temp_regs[decl->Range.First + i].offset = i;
> + }
> + } else {
> + for (i = 0; i < array_size; i++) {
> + reg = nir_local_reg_create(b->impl);
> + reg->num_components = 4;
> + c->temp_regs[decl->Range.First + i].reg = reg;
> + c->temp_regs[decl->Range.First + i].offset = 0;
> + }
> + }
> + } else if (file == TGSI_FILE_ADDRESS) {
> + c->addr_reg = nir_local_reg_create(b->impl);
> + c->addr_reg->num_components = 4;
> + } else if (file == TGSI_FILE_SAMPLER) {
> + /* Nothing to record for samplers. */
> + } else {
> + nir_variable *var;
> + assert(file == TGSI_FILE_INPUT ||
> + file == TGSI_FILE_OUTPUT ||
> + file == TGSI_FILE_CONSTANT);
> +
> + var = rzalloc(b->shader, nir_variable);
> + var->data.driver_location = decl->Range.First;
> +
> + var->type = glsl_vec4_type();
> + if (array_size > 1)
> + var->type = glsl_array_type(var->type, array_size);
> +
> + switch (file) {
> + case TGSI_FILE_INPUT:
> + var->data.read_only = true;
> + var->data.mode = nir_var_shader_in;
> + var->name = ralloc_asprintf(var, "in_%d", decl->Range.First);
> +
> + /* We should probably translate to a VERT_ATTRIB_* or
VARYING_SLOT_*
> + * instead, but nothing in NIR core is looking at the value
> + * currently, and this is less change to drivers.
> + */
> + var->data.location = decl->Semantic.Name;
> + var->data.index = decl->Semantic.Index;
> +
> + /* We definitely need to translate the interpolation field, because
> + * nir_print will decode it.
> + */
> + switch (decl->Interp.Interpolate) {
> + case TGSI_INTERPOLATE_CONSTANT:
> + var->data.interpolation = INTERP_QUALIFIER_FLAT;
> + break;
> + case TGSI_INTERPOLATE_LINEAR:
> + var->data.interpolation = INTERP_QUALIFIER_NOPERSPECTIVE;
> + break;
> + case TGSI_INTERPOLATE_PERSPECTIVE:
> + var->data.interpolation = INTERP_QUALIFIER_SMOOTH;
> + break;
> + }
> +
> + exec_list_push_tail(&b->shader->inputs, &var->node);
> + break;
> + case TGSI_FILE_OUTPUT: {
> + /* Since we can't load from outputs in the IR, we make temporaries
> + * for the outputs and emit stores to the real outputs at the end
of
> + * the shader.
> + */
> + nir_register *reg = nir_local_reg_create(b->impl);
> + reg->num_components = 4;
> + if (array_size > 1)
> + reg->num_array_elems = array_size;
> +
> + var->data.mode = nir_var_shader_out;
> + var->name = ralloc_asprintf(var, "out_%d", decl->Range.First);
> +
> + var->data.location = decl->Semantic.Name;
> + var->data.index = decl->Semantic.Index;
> +
> + for (i = 0; i < array_size; i++) {
> + c->output_regs[decl->Range.First + i].offset = i;
> + c->output_regs[decl->Range.First + i].reg = reg;
> + }
> +
> + exec_list_push_tail(&b->shader->outputs, &var->node);
> + }
> + break;
> + case TGSI_FILE_CONSTANT:
> + var->data.mode = nir_var_uniform;
> + var->name = ralloc_asprintf(var, "uniform_%d", decl->Range.First);
> +
> + exec_list_push_tail(&b->shader->uniforms, &var->node);
So...here, you add a bunch of variables to shader->uniforms, but you
never actually refer to those variables in this pass - you emit
intrinsics to load uniforms at a particular index directly. Seems odd.
I see ntq_setup_uniforms() uses them, though, so perhaps they're still
useful. You don't appear to be setting shader->num_uniforms, which
might be worth doing.
In i965/glsl, glsl_to_nir creates variables in shader->uniforms, and
nir_assign_var_locations_scalar_direct_first sets shader->num_uniforms
and lowers the variable accesses to uniform load intrinsics with indices.
In i965/prog, prog_to_nir sets shader->num_uniforms, but leaves the list
of uniform variables blank. It generates load intrinsics, and the
backend just sets up uniforms based on the prog->Parameters list.
Given that TGSI gives you actual information about variables (instead of
a giant array like Mesa IR), this seems like a reasonable approach.
> + break;
> + default:
> + unreachable("bad declaration file");
> + return;
> + }
> +
> + }
> +}
> +
> +static void
> +ttn_emit_immediate(struct ttn_compile *c)
> +{
> + nir_builder *b = &c->build;
> + struct tgsi_full_immediate *tgsi_imm = &c->token->FullImmediate;
> + nir_load_const_instr *load_const;
> + int i;
> +
> + load_const = nir_load_const_instr_create(b->shader, 4);
> + c->imm_defs[c->next_imm] = &load_const->def;
> + c->next_imm++;
> +
> + for (i = 0; i < 4; i++)
> + load_const->value.u[i] = tgsi_imm->u[i].Uint;
> +
> + nir_instr_insert_after_cf_list(b->cf_node_list, &load_const->instr);
> +}
> +
> +static nir_src *
> +ttn_src_for_indirect(struct ttn_compile *c, struct tgsi_ind_register
*indirect);
> +
> +static nir_src
> +ttn_src_for_file_and_index(struct ttn_compile *c, unsigned file, unsigned
index,
> + struct tgsi_ind_register *indirect)
> +{
> + nir_builder *b = &c->build;
> + nir_src src;
> +
> + memset(&src, 0, sizeof(src));
> +
> + switch (file) {
> + case TGSI_FILE_TEMPORARY:
> + src.reg.reg = c->temp_regs[index].reg;
> + src.reg.base_offset = c->temp_regs[index].offset;
> + if (indirect)
> + src.reg.indirect = ttn_src_for_indirect(c, indirect);
> + break;
> +
> + case TGSI_FILE_ADDRESS:
> + src.reg.reg = c->addr_reg;
> + break;
> +
> + case TGSI_FILE_IMMEDIATE:
> + src = nir_src_for_ssa(c->imm_defs[index]);
> + assert(!indirect);
> + break;
> +
> + case TGSI_FILE_INPUT:
> + case TGSI_FILE_CONSTANT: {
> + nir_intrinsic_instr *load;
> +
> + switch (file) {
> + case TGSI_FILE_INPUT:
> + load = nir_intrinsic_instr_create(b->shader,
> + indirect ?
> +
nir_intrinsic_load_input_indirect :
> + nir_intrinsic_load_input);
> + break;
Huh. Kudos on getting this working, I never managed. I read the input
variables directly and relied on nir_lower_io to sort it out for me.
Seems reasonable.
> + case TGSI_FILE_CONSTANT:
> + load = nir_intrinsic_instr_create(b->shader,
> + indirect ?
> +
nir_intrinsic_load_uniform_indirect :
> + nir_intrinsic_load_uniform);
> + break;
> + default:
> + unreachable("No other load files supported");
> + break;
> + }
> +
> + load->num_components = 4;
> + load->const_index[0] = index;
> + load->const_index[1] = 1;
> + if (indirect) {
> + nir_alu_src indirect_address;
> + memset(&indirect_address, 0, sizeof(indirect_address));
> + indirect_address.src = nir_src_for_reg(c->addr_reg);
> + for (int i = 0; i < 4; i++)
> + indirect_address.swizzle[i] = indirect->Swizzle;
> + load->src[0] = nir_src_for_ssa(nir_imov_alu(b, indirect_address,
4));
> + }
> + nir_ssa_dest_init(&load->instr, &load->dest, 4, NULL);
> + nir_instr_insert_after_cf_list(b->cf_node_list, &load->instr);
> +
> + src = nir_src_for_ssa(&load->dest.ssa);
> + break;
> + }
> +
> + default:
> + unreachable("bad src file");
> + }
> +
> +
> + return src;
> +}
> +
> +static nir_src *
> +ttn_src_for_indirect(struct ttn_compile *c, struct tgsi_ind_register
*indirect)
> +{
> + nir_builder *b = &c->build;
> + nir_alu_src src;
> + memset(&src, 0, sizeof(src));
> + for (int i = 0; i < 4; i++)
> + src.swizzle[i] = indirect->Swizzle;
> + src.src = ttn_src_for_file_and_index(c,
> + indirect->File,
> + indirect->Index, NULL);
> + nir_src *result = ralloc(b->shader, nir_src);
> + *result = nir_src_for_ssa(nir_imov_alu(b, src, 1));
> + return result;
> +}
> +
> +static nir_alu_dest
> +ttn_get_dest(struct ttn_compile *c, struct tgsi_full_dst_register
*tgsi_fdst)
> +{
> + struct tgsi_dst_register *tgsi_dst = &tgsi_fdst->Register;
> + nir_alu_dest dest;
> +
> + memset(&dest, 0, sizeof(dest));
> +
> + if (tgsi_dst->File == TGSI_FILE_TEMPORARY) {
> + dest.dest.reg.reg = c->temp_regs[tgsi_dst->Index].reg;
> + dest.dest.reg.base_offset = c->temp_regs[tgsi_dst->Index].offset;
> + } else if (tgsi_dst->File == TGSI_FILE_OUTPUT) {
> + dest.dest.reg.reg = c->output_regs[tgsi_dst->Index].reg;
> + dest.dest.reg.base_offset = c->output_regs[tgsi_dst->Index].offset;
> + } else if (tgsi_dst->File == TGSI_FILE_ADDRESS) {
> + assert(tgsi_dst->Index == 0);
> + dest.dest.reg.reg = c->addr_reg;
> + }
> +
> + dest.write_mask = tgsi_dst->WriteMask;
> + dest.saturate = false;
> +
> + if (tgsi_dst->Indirect)
> + dest.dest.reg.indirect = ttn_src_for_indirect(c, &tgsi_fdst-
>Indirect);
> +
> + return dest;
> +}
> +
> +static nir_ssa_def *
> +ttn_get_src(struct ttn_compile *c, struct tgsi_full_src_register
*tgsi_fsrc)
> +{
> + nir_builder *b = &c->build;
> + struct tgsi_src_register *tgsi_src = &tgsi_fsrc->Register;
> + unsigned tgsi_opcode = c->token->FullInstruction.Instruction.Opcode;
> + unsigned tgsi_src_type = tgsi_opcode_infer_src_type(tgsi_opcode);
> + bool src_is_float = !(tgsi_src_type == TGSI_TYPE_SIGNED ||
> + tgsi_src_type == TGSI_TYPE_UNSIGNED);
> + nir_alu_src src;
> +
> + memset(&src, 0, sizeof(src));
> +
> + if (tgsi_src->File == TGSI_FILE_NULL) {
> + return nir_imm_float(b, 0.0);
> + } else if (tgsi_src->File == TGSI_FILE_SAMPLER) {
> + /* Only the index of the sampler gets used in texturing, and it will
> + * handle looking that up on its own instead of using the
nir_alu_src.
> + */
> + assert(!tgsi_src->Indirect);
> + return NULL;
> + } else {
> + src.src = ttn_src_for_file_and_index(c,
> + tgsi_src->File,
> + tgsi_src->Index,
> + (tgsi_src->Indirect ?
> + &tgsi_fsrc->Indirect : NULL));
> + }
> +
> + src.swizzle[0] = tgsi_src->SwizzleX;
> + src.swizzle[1] = tgsi_src->SwizzleY;
> + src.swizzle[2] = tgsi_src->SwizzleZ;
> + src.swizzle[3] = tgsi_src->SwizzleW;
> +
> + nir_ssa_def *def = nir_fmov_alu(b, src, 4);
> +
> + if (tgsi_src->Absolute) {
> + if (src_is_float)
> + def = nir_fabs(b, def);
> + else
> + def = nir_iabs(b, def);
> + }
> +
> + if (tgsi_src->Negate) {
> + if (src_is_float)
> + def = nir_fneg(b, def);
> + else
> + def = nir_ineg(b, def);
> + }
> +
> + return def;
> +}
> +
> +static void
> +ttn_alu(nir_builder *b, nir_op op, nir_alu_dest dest, nir_ssa_def **src)
> +{
> + unsigned num_srcs = nir_op_infos[op].num_inputs;
> + nir_alu_instr *instr = nir_alu_instr_create(b->shader, op);
> + unsigned i;
> +
> + for (i = 0; i < num_srcs; i++)
> + instr->src[i].src = nir_src_for_ssa(src[i]);
> +
> + instr->dest = dest;
> + nir_instr_insert_after_cf_list(b->cf_node_list, &instr->instr);
> +}
> +
> +static void
> +ttn_move_dest_masked(nir_builder *b, nir_alu_dest dest,
> + nir_ssa_def *def, unsigned write_mask)
> +{
> + if (!(dest.write_mask & write_mask))
> + return;
> +
> + nir_alu_instr *mov = nir_alu_instr_create(b->shader, nir_op_imov);
> + mov->dest = dest;
> + mov->dest.write_mask &= write_mask;
> + mov->src[0].src = nir_src_for_ssa(def);
> + for (unsigned i = def->num_components; i < 4; i++)
> + mov->src[0].swizzle[i] = def->num_components - 1;
> + nir_instr_insert_after_cf_list(b->cf_node_list, &mov->instr);
> +}
> +
> +static void
> +ttn_move_dest(nir_builder *b, nir_alu_dest dest, nir_ssa_def *def)
> +{
> + ttn_move_dest_masked(b, dest, def, TGSI_WRITEMASK_XYZW);
> +}
> +
> +static void
> +ttn_arl(nir_builder *b, nir_op op, nir_alu_dest dest, nir_ssa_def **src)
> +{
> + ttn_move_dest(b, dest, nir_f2i(b, nir_ffloor(b, src[0])));
> +}
> +
> +/* EXP - Approximate Exponential Base 2
> + * dst.x = 2^{\lfloor src.x\rfloor}
> + * dst.y = src.x - \lfloor src.x\rfloor
> + * dst.z = 2^{src.x}
> + * dst.w = 1.0
> + */
> +static void
> +ttn_exp(nir_builder *b, nir_op op, nir_alu_dest dest, nir_ssa_def **src)
> +{
> + nir_ssa_def *srcx = ttn_channel(b, src[0], X);
> +
> + ttn_move_dest_masked(b, dest, nir_fexp2(b, nir_ffloor(b, srcx)),
> + TGSI_WRITEMASK_X);
> + ttn_move_dest_masked(b, dest, nir_fsub(b, srcx, nir_ffloor(b, srcx)),
> + TGSI_WRITEMASK_Y);
> + ttn_move_dest_masked(b, dest, nir_fexp2(b, srcx), TGSI_WRITEMASK_Z);
> + ttn_move_dest_masked(b, dest, nir_imm_float(b, 1.0), TGSI_WRITEMASK_W);
> +}
> +
> +/* LOG - Approximate Logarithm Base 2
> + * dst.x = \lfloor\log_2{|src.x|}\rfloor
> + * dst.y = \frac{|src.x|}{2^{\lfloor\log_2{|src.x|}\rfloor}}
> + * dst.z = \log_2{|src.x|}
> + * dst.w = 1.0
> + */
> +static void
> +ttn_log(nir_builder *b, nir_op op, nir_alu_dest dest, nir_ssa_def **src)
> +{
> + nir_ssa_def *abs_srcx = nir_fabs(b, ttn_channel(b, src[0], X));
> + nir_ssa_def *log2 = nir_flog2(b, abs_srcx);
> +
> + ttn_move_dest_masked(b, dest, nir_ffloor(b, log2), TGSI_WRITEMASK_X);
> + ttn_move_dest_masked(b, dest,
> + nir_fdiv(b, abs_srcx, nir_fexp2(b, nir_ffloor(b,
log2))),
You're generating two copies of floor(log2) here, which will have to be
CSE'd later. In prog_to_nir, I created a temporary and used it in both
places:
nir_ssa_def *floor_log2 = nir_ffloor(b, log2);
We're generating tons of rubbish for NIR to optimize anyway, so it's not
a big deal...but...may as well do the trivial improvement.
> + TGSI_WRITEMASK_Y);
> + ttn_move_dest_masked(b, dest, nir_flog2(b, abs_srcx), TGSI_WRITEMASK_Z);
> + ttn_move_dest_masked(b, dest, nir_imm_float(b, 1.0), TGSI_WRITEMASK_W);
> +}
> +
> +/* DST - Distance Vector
> + * dst.x = 1.0
> + * dst.y = src0.y \times src1.y
> + * dst.z = src0.z
> + * dst.w = src1.w
> + */
> +static void
> +ttn_dst(nir_builder *b, nir_op op, nir_alu_dest dest, nir_ssa_def **src)
> +{
> + ttn_move_dest_masked(b, dest, nir_imm_float(b, 1.0), TGSI_WRITEMASK_X);
> + ttn_move_dest_masked(b, dest, nir_fmul(b, src[0], src[1]),
TGSI_WRITEMASK_Y);
> + ttn_move_dest_masked(b, dest, nir_fmov(b, src[0]), TGSI_WRITEMASK_Z);
> + ttn_move_dest_masked(b, dest, nir_fmov(b, src[1]), TGSI_WRITEMASK_W);
> +}
> +
> +/* LIT - Light Coefficients
> + * dst.x = 1.0
> + * dst.y = max(src.x, 0.0)
> + * dst.z = (src.x > 0.0) ? max(src.y, 0.0)^{clamp(src.w, -128.0, 128.0))}
: 0
> + * dst.w = 1.0
> + */
> +static void
> +ttn_lit(nir_builder *b, nir_op op, nir_alu_dest dest, nir_ssa_def **src)
> +{
> + ttn_move_dest_masked(b, dest, nir_imm_float(b, 1.0), TGSI_WRITEMASK_XW);
> +
> + ttn_move_dest_masked(b, dest, nir_fmax(b, ttn_channel(b, src[0], X),
> + nir_imm_float(b, 0.0)),
TGSI_WRITEMASK_Y);
> +
> + if (dest.write_mask & TGSI_WRITEMASK_Z) {
> + nir_ssa_def *src0_y = ttn_channel(b, src[0], Y);
> + nir_ssa_def *wclamp = nir_fmax(b, nir_fmin(b, ttn_channel(b, src[0],
W),
> + nir_imm_float(b, 128.0)),
> + nir_imm_float(b, -128.0));
> + nir_ssa_def *pow = nir_fpow(b, nir_fmax(b, src0_y, nir_imm_float(b,
0.0)),
> + wclamp);
> +
> + ttn_move_dest_masked(b, dest,
> + nir_bcsel(b,
> + nir_fge(b,
> + nir_imm_float(b, 0.0),
> + ttn_channel(b, src[0], X)),
> + nir_imm_float(b, 0.0),
> + pow),
> + TGSI_WRITEMASK_Z);
> + }
> +}
> +
> +/* SCS - Sine Cosine
> + * dst.x = \cos{src.x}
> + * dst.y = \sin{src.x}
> + * dst.z = 0.0
> + * dst.w = 1.0
> + */
> +static void
> +ttn_scs(nir_builder *b, nir_op op, nir_alu_dest dest, nir_ssa_def **src)
> +{
> + ttn_move_dest_masked(b, dest, nir_fcos(b, ttn_channel(b, src[0], X)),
> + TGSI_WRITEMASK_X);
> + ttn_move_dest_masked(b, dest, nir_fsin(b, ttn_channel(b, src[0], X)),
> + TGSI_WRITEMASK_Y);
> + ttn_move_dest_masked(b, dest, nir_imm_float(b, 0.0), TGSI_WRITEMASK_Z);
> + ttn_move_dest_masked(b, dest, nir_imm_float(b, 1.0), TGSI_WRITEMASK_W);
> +}
> +
> +static void
> +ttn_sle(nir_builder *b, nir_op op, nir_alu_dest dest, nir_ssa_def **src)
> +{
> + ttn_move_dest(b, dest, nir_sge(b, src[1], src[0]));
> +}
I've got code here to generate b2f(fge(...)) instead of sge(...) since I
didn't want to bother implementing it in my driver, and figured the b2fs
might be able to get optimized away.
That said, I suppose we could probably just add lowering transformations
that turn sge -> b2f(fge(...)) when options->native_integers is set, and
delete my code...
> +static void
> +ttn_sgt(nir_builder *b, nir_op op, nir_alu_dest dest, nir_ssa_def **src)
> +{
> + ttn_move_dest(b, dest, nir_slt(b, src[1], src[0]));
> +}
> +
> +static void
> +ttn_clamp(nir_builder *b, nir_op op, nir_alu_dest dest, nir_ssa_def **src)
> +{
> + ttn_move_dest(b, dest, nir_fmin(b, nir_fmax(b, src[0], src[1]),
src[2]));
> +}
> +
> +static void
> +ttn_xpd(nir_builder *b, nir_op op, nir_alu_dest dest, nir_ssa_def **src)
> +{
> + ttn_move_dest_masked(b, dest,
> + nir_fsub(b,
> + nir_fmul(b,
> + ttn_swizzle(b, src[0], Y, Z, X,
X),
> + ttn_swizzle(b, src[1], Z, X, Y,
X)),
> + nir_fmul(b,
> + ttn_swizzle(b, src[1], Y, Z, X,
X),
> + ttn_swizzle(b, src[0], Z, X, Y,
X))),
> + TGSI_WRITEMASK_XYZ);
> + ttn_move_dest_masked(b, dest, nir_imm_float(b, 1.0), TGSI_WRITEMASK_W);
> +}
> +
> +static void
> +ttn_dp2a(nir_builder *b, nir_op op, nir_alu_dest dest, nir_ssa_def **src)
> +{
> + ttn_move_dest(b, dest,
> + ttn_channel(b, nir_fadd(b,
> + ttn_channel(b, nir_fdot2(b,
src[0],
> + src[1]),
> + X),
Do you really need to do ttn_channel(b, ..., X) on a fdot2 result? It's
already a scalar value. Same comment applies to the below four.
I should probably delete that from prog_to_nir as well.
> + src[2]),
> + X));
> +}
> +
> +static void
> +ttn_dp2(nir_builder *b, nir_op op, nir_alu_dest dest, nir_ssa_def **src)
> +{
> + ttn_move_dest(b, dest, ttn_channel(b, nir_fdot2(b, src[0], src[1]), X));
> +}
> +
> +static void
> +ttn_dp3(nir_builder *b, nir_op op, nir_alu_dest dest, nir_ssa_def **src)
> +{
> + ttn_move_dest(b, dest, ttn_channel(b, nir_fdot3(b, src[0], src[1]), X));
> +}
> +
> +static void
> +ttn_dp4(nir_builder *b, nir_op op, nir_alu_dest dest, nir_ssa_def **src)
> +{
> + ttn_move_dest(b, dest, ttn_channel(b, nir_fdot4(b, src[0], src[1]), X));
> +}
> +
> +static void
> +ttn_dph(nir_builder *b, nir_op op, nir_alu_dest dest, nir_ssa_def **src)
> +{
> + nir_ssa_def *dp3 = ttn_channel(b, nir_fdot3(b, src[0], src[1]), X);
> + ttn_move_dest(b, dest, nir_fadd(b, dp3, ttn_channel(b, src[1], W)));
> +}
> +
> +static void
> +ttn_arr(nir_builder *b, nir_op op, nir_alu_dest dest, nir_ssa_def **src)
> +{
> + ttn_move_dest(b, dest, nir_ffloor(b, nir_fadd(b, src[0], nir_imm_float(b,
0.5))));
> +}
> +
> +static void
> +ttn_cmp(nir_builder *b, nir_op op, nir_alu_dest dest, nir_ssa_def **src)
> +{
> + ttn_move_dest(b, dest, nir_bcsel(b,
> + nir_flt(b, src[0], nir_imm_float(b,
0.0)),
> + src[1], src[2]));
> +}
> +
> +static void
> +ttn_ucmp(nir_builder *b, nir_op op, nir_alu_dest dest, nir_ssa_def **src)
> +{
> + ttn_move_dest(b, dest, nir_bcsel(b,
> + nir_ine(b, src[0], nir_imm_float(b,
0.0)),
Doing nir_imm_int(b, 0) here would make more sense.
> + src[1], src[2]));
> +}
> +
> +static void
> +ttn_kill(nir_builder *b, nir_op op, nir_alu_dest dest, nir_ssa_def **src)
> +{
> + nir_intrinsic_instr *discard =
> + nir_intrinsic_instr_create(b->shader, nir_intrinsic_discard);
> + nir_instr_insert_after_cf_list(b->cf_node_list, &discard->instr);
> +}
> +
> +static void
> +ttn_kill_if(nir_builder *b, nir_op op, nir_alu_dest dest, nir_ssa_def
**src)
> +{
> + nir_if *if_stmt = nir_if_create(b->shader);
> + if_stmt->condition =
> + nir_src_for_ssa(nir_bany4(b, nir_flt(b, src[0], nir_imm_float(b,
0.0))));
> + nir_cf_node_insert_end(b->cf_node_list, &if_stmt->cf_node);
> +
> + nir_intrinsic_instr *discard =
> + nir_intrinsic_instr_create(b->shader, nir_intrinsic_discard);
> + nir_instr_insert_after_cf_list(&if_stmt->then_list, &discard->instr);
A while back I created a discard_if intrinsic for exactly this case.
You just want:
nir_ssa_def *cmp = nir_bany4(b, nir_flt(b, src[0], nir_imm_float(b, 0.0)));
nir_intrinsic_instr *discard =
nir_intrinsic_instr_create(b->shader, nir_intrinsic_discard_if);
discard->src[0] = nir_src_for_ssa(cmp);
nir_instr_insert_after_cf_list(b->cf_node_list, &discard->instr);
> +}
> +
> +static void
> +ttn_if(struct ttn_compile *c, nir_ssa_def *src, bool is_uint)
> +{
> + nir_builder *b = &c->build;
> +
> + /* Save the outside-of-the-if-statement node list. */
> + c->if_stack[c->if_stack_pos] = b->cf_node_list;
> + c->if_stack_pos++;
> +
> + src = ttn_channel(b, src, X);
> +
> + nir_if *if_stmt = nir_if_create(b->shader);
> + if (is_uint) {
> + if_stmt->condition = nir_src_for_ssa(nir_ine(b, src, nir_imm_int(b,
0)));
> + } else {
> + if_stmt->condition = nir_src_for_ssa(nir_fne(b, src, nir_imm_int(b,
0)));
> + }
> + nir_cf_node_insert_end(b->cf_node_list, &if_stmt->cf_node);
> +
> + nir_builder_insert_after_cf_list(b, &if_stmt->then_list);
> +
> + c->if_stack[c->if_stack_pos] = &if_stmt->else_list;
> + c->if_stack_pos++;
> +}
> +
> +static void
> +ttn_else(struct ttn_compile *c)
> +{
> + nir_builder *b = &c->build;
> +
> + nir_builder_insert_after_cf_list(b, c->if_stack[c->if_stack_pos - 1]);
> +}
> +
> +static void
> +ttn_endif(struct ttn_compile *c)
> +{
> + nir_builder *b = &c->build;
> +
> + c->if_stack_pos -= 2;
Why 2? I don't see ELSE or ENDIF being put on the stack, so aren't you
just trying to pop off the single IF?
> + nir_builder_insert_after_cf_list(b, c->if_stack[c->if_stack_pos]);
> +}
> +
> +static void
> +ttn_bgnloop(struct ttn_compile *c)
> +{
> + nir_builder *b = &c->build;
> +
> + /* Save the outside-of-the-loop node list. */
> + c->loop_stack[c->loop_stack_pos] = b->cf_node_list;
> + c->loop_stack_pos++;
> +
> + nir_loop *loop = nir_loop_create(b->shader);
> + nir_cf_node_insert_end(b->cf_node_list, &loop->cf_node);
> +
> + nir_builder_insert_after_cf_list(b, &loop->body);
> +}
> +
> +static void
> +ttn_cont(nir_builder *b)
> +{
> + nir_jump_instr *instr = nir_jump_instr_create(b->shader,
nir_jump_continue);
> + nir_instr_insert_after_cf_list(b->cf_node_list, &instr->instr);
> +}
> +
> +static void
> +ttn_brk(nir_builder *b)
> +{
> + nir_jump_instr *instr = nir_jump_instr_create(b->shader,
nir_jump_break);
> + nir_instr_insert_after_cf_list(b->cf_node_list, &instr->instr);
> +}
> +
> +static void
> +ttn_endloop(struct ttn_compile *c)
> +{
> + nir_builder *b = &c->build;
> +
> + c->loop_stack_pos--;
> + nir_builder_insert_after_cf_list(b, c->loop_stack[c->loop_stack_pos]);
> +}
> +
> +static void
> +ttn_tex(struct ttn_compile *c, nir_alu_dest dest, nir_ssa_def **src)
> +{
> + nir_builder *b = &c->build;
> + struct tgsi_full_instruction *tgsi_inst = &c->token->FullInstruction;
> + nir_tex_instr *instr;
> + nir_texop op;
> + unsigned num_srcs;
> +
> + switch (tgsi_inst->Instruction.Opcode) {
> + case TGSI_OPCODE_TEX:
> + op = nir_texop_tex;
> + num_srcs = 1;
> + break;
> + case TGSI_OPCODE_TXP:
> + op = nir_texop_tex;
> + num_srcs = 2;
> + break;
> + case TGSI_OPCODE_TXB:
> + op = nir_texop_txb;
> + num_srcs = 2;
> + break;
> + case TGSI_OPCODE_TXL:
> + op = nir_texop_txl;
> + num_srcs = 2;
> + break;
> + case TGSI_OPCODE_TXF:
> + op = nir_texop_txf;
> + num_srcs = 1;
> + break;
> + case TGSI_OPCODE_TXD:
> + op = nir_texop_txd;
> + num_srcs = 3;
> + break;
> + default:
> + fprintf(stderr, "unknown TGSI tex op %d\n", tgsi_inst-
>Instruction.Opcode);
> + abort();
> + }
> +
> + if (tgsi_inst->Texture.Texture == TGSI_TEXTURE_SHADOW1D ||
> + tgsi_inst->Texture.Texture == TGSI_TEXTURE_SHADOW1D_ARRAY ||
> + tgsi_inst->Texture.Texture == TGSI_TEXTURE_SHADOW2D ||
> + tgsi_inst->Texture.Texture == TGSI_TEXTURE_SHADOW2D_ARRAY ||
> + tgsi_inst->Texture.Texture == TGSI_TEXTURE_SHADOWRECT ||
> + tgsi_inst->Texture.Texture == TGSI_TEXTURE_SHADOWCUBE ||
> + tgsi_inst->Texture.Texture == TGSI_TEXTURE_SHADOWCUBE_ARRAY) {
> + num_srcs++;
> + }
> +
> + instr = nir_tex_instr_create(b->shader, num_srcs);
> + instr->op = op;
> +
> + switch (tgsi_inst->Texture.Texture) {
> + case TGSI_TEXTURE_1D:
> + instr->sampler_dim = GLSL_SAMPLER_DIM_1D;
> + break;
> + case TGSI_TEXTURE_1D_ARRAY:
> + instr->sampler_dim = GLSL_SAMPLER_DIM_1D;
> + instr->is_array = true;
> + break;
> + case TGSI_TEXTURE_SHADOW1D:
> + instr->sampler_dim = GLSL_SAMPLER_DIM_1D;
> + instr->is_shadow = true;
> + break;
> + case TGSI_TEXTURE_SHADOW1D_ARRAY:
> + instr->sampler_dim = GLSL_SAMPLER_DIM_1D;
> + instr->is_shadow = true;
> + instr->is_array = true;
> + break;
> + case TGSI_TEXTURE_2D:
> + instr->sampler_dim = GLSL_SAMPLER_DIM_2D;
> + break;
> + case TGSI_TEXTURE_2D_ARRAY:
> + instr->sampler_dim = GLSL_SAMPLER_DIM_2D;
> + instr->is_array = true;
> + break;
> + case TGSI_TEXTURE_2D_MSAA:
> + instr->sampler_dim = GLSL_SAMPLER_DIM_MS;
> + break;
> + case TGSI_TEXTURE_2D_ARRAY_MSAA:
> + instr->sampler_dim = GLSL_SAMPLER_DIM_MS;
> + instr->is_array = true;
> + break;
> + case TGSI_TEXTURE_SHADOW2D:
> + instr->sampler_dim = GLSL_SAMPLER_DIM_2D;
> + instr->is_shadow = true;
> + break;
> + case TGSI_TEXTURE_SHADOW2D_ARRAY:
> + instr->sampler_dim = GLSL_SAMPLER_DIM_2D;
> + instr->is_shadow = true;
> + instr->is_array = true;
> + break;
> + case TGSI_TEXTURE_3D:
> + instr->sampler_dim = GLSL_SAMPLER_DIM_3D;
> + break;
> + case TGSI_TEXTURE_CUBE:
> + instr->sampler_dim = GLSL_SAMPLER_DIM_CUBE;
> + break;
> + case TGSI_TEXTURE_CUBE_ARRAY:
> + instr->sampler_dim = GLSL_SAMPLER_DIM_CUBE;
> + instr->is_array = true;
> + break;
FWIW, I don't think CUBE_ARRAY and SHADOWCUBE_ARRAY will work with
this pass...I'm pretty sure you need to implement TXB2/TXL2. You call
this function for those, but they fall through to "unknown TGSI tex op".
I suspect you don't care for your hardware, so punting and making
someone else do it seems reasonable. :)
> + case TGSI_TEXTURE_SHADOWCUBE:
> + instr->sampler_dim = GLSL_SAMPLER_DIM_CUBE;
> + instr->is_shadow = true;
> + break;
> + case TGSI_TEXTURE_SHADOWCUBE_ARRAY:
> + instr->sampler_dim = GLSL_SAMPLER_DIM_CUBE;
> + instr->is_shadow = true;
> + instr->is_array = true;
> + break;
> + case TGSI_TEXTURE_RECT:
> + instr->sampler_dim = GLSL_SAMPLER_DIM_RECT;
> + break;
> + case TGSI_TEXTURE_SHADOWRECT:
> + instr->sampler_dim = GLSL_SAMPLER_DIM_RECT;
> + instr->is_shadow = true;
> + break;
> + default:
> + fprintf(stderr, "Unknown TGSI texture target %d\n",
> + tgsi_inst->Texture.Texture);
> + abort();
> + }
> +
> + switch (instr->sampler_dim) {
> + case GLSL_SAMPLER_DIM_1D:
> + case GLSL_SAMPLER_DIM_BUF:
> + instr->coord_components = 1;
> + break;
> + case GLSL_SAMPLER_DIM_2D:
> + case GLSL_SAMPLER_DIM_RECT:
> + case GLSL_SAMPLER_DIM_EXTERNAL:
> + case GLSL_SAMPLER_DIM_MS:
> + instr->coord_components = 2;
> + break;
> + case GLSL_SAMPLER_DIM_3D:
> + case GLSL_SAMPLER_DIM_CUBE:
> + instr->coord_components = 3;
> + break;
> + }
> +
> + if (instr->is_array)
> + instr->coord_components++;
> +
> + assert(tgsi_inst->Src[1].Register.File == TGSI_FILE_SAMPLER);
> + instr->sampler_index = tgsi_inst->Src[1].Register.Index;
> +
> + unsigned src_number = 0;
> +
> + if (tgsi_inst->Instruction.Opcode != TGSI_OPCODE_TXQ) {
> + instr->src[src_number].src =
> + nir_src_for_ssa(nir_swizzle(b, src[0], SWIZ(X, Y, Z, W),
> + instr->coord_components, false));
> + instr->src[src_number].src_type = nir_tex_src_coord;
> + src_number++;
> + }
> +
> + if (tgsi_inst->Instruction.Opcode == TGSI_OPCODE_TXP) {
> + instr->src[src_number].src = nir_src_for_ssa(ttn_channel(b, src[0],
W));
> + instr->src[src_number].src_type = nir_tex_src_projector;
> + src_number++;
> + }
> +
> + if (tgsi_inst->Instruction.Opcode == TGSI_OPCODE_TXB) {
> + instr->src[src_number].src = nir_src_for_ssa(ttn_channel(b, src[0],
W));
> + instr->src[src_number].src_type = nir_tex_src_bias;
> + src_number++;
> + }
> +
> + if (tgsi_inst->Instruction.Opcode == TGSI_OPCODE_TXL) {
> + instr->src[src_number].src = nir_src_for_ssa(ttn_channel(b, src[0],
W));
> + instr->src[src_number].src_type = nir_tex_src_lod;
> + src_number++;
> + }
> +
> + if (instr->is_shadow) {
> + if (instr->coord_components < 3)
> + instr->src[src_number].src = nir_src_for_ssa(ttn_channel(b,
src[0], Z));
> + else
> + instr->src[src_number].src = nir_src_for_ssa(ttn_channel(b,
src[0], W));
> +
> + instr->src[src_number].src_type = nir_tex_src_comparitor;
> + src_number++;
> + }
> +
> + assert(src_number == num_srcs);
> +
> + nir_ssa_dest_init(&instr->instr, &instr->dest, 4, NULL);
> + nir_instr_insert_after_cf_list(b->cf_node_list, &instr->instr);
> +
> + /* Resolve the writemask on the texture op. */
> + ttn_move_dest(b, dest, &instr->dest.ssa);
> +}
> +
> +static const nir_op op_trans[TGSI_OPCODE_LAST] = {
> + [TGSI_OPCODE_ARL] = 0,
> + [TGSI_OPCODE_MOV] = nir_op_fmov,
> + [TGSI_OPCODE_LIT] = 0,
> + [TGSI_OPCODE_RCP] = nir_op_frcp,
> + [TGSI_OPCODE_RSQ] = nir_op_frsq,
> + [TGSI_OPCODE_EXP] = 0,
> + [TGSI_OPCODE_LOG] = 0,
> + [TGSI_OPCODE_MUL] = nir_op_fmul,
> + [TGSI_OPCODE_ADD] = nir_op_fadd,
> + [TGSI_OPCODE_DP3] = 0,
> + [TGSI_OPCODE_DP4] = 0,
> + [TGSI_OPCODE_DST] = 0,
> + [TGSI_OPCODE_MIN] = nir_op_fmin,
> + [TGSI_OPCODE_MAX] = nir_op_fmax,
> + [TGSI_OPCODE_SLT] = nir_op_slt,
> + [TGSI_OPCODE_SGE] = nir_op_sge,
> + [TGSI_OPCODE_MAD] = nir_op_ffma,
> + [TGSI_OPCODE_SUB] = nir_op_fsub,
> + [TGSI_OPCODE_LRP] = 0,
> + [TGSI_OPCODE_SQRT] = nir_op_fsqrt,
> + [TGSI_OPCODE_DP2A] = 0,
> + [TGSI_OPCODE_FRC] = nir_op_ffract,
> + [TGSI_OPCODE_CLAMP] = 0,
> + [TGSI_OPCODE_FLR] = nir_op_ffloor,
> + [TGSI_OPCODE_ROUND] = nir_op_fround_even,
> + [TGSI_OPCODE_EX2] = nir_op_fexp2,
> + [TGSI_OPCODE_LG2] = nir_op_flog2,
> + [TGSI_OPCODE_POW] = nir_op_fpow,
> + [TGSI_OPCODE_XPD] = 0,
> + [TGSI_OPCODE_ABS] = nir_op_fabs,
> + [TGSI_OPCODE_DPH] = 0,
> + [TGSI_OPCODE_COS] = nir_op_fcos,
> + [TGSI_OPCODE_DDX] = nir_op_fddx,
> + [TGSI_OPCODE_DDY] = nir_op_fddy,
> + [TGSI_OPCODE_KILL] = 0,
> + [TGSI_OPCODE_PK2H] = 0, /* XXX */
> + [TGSI_OPCODE_PK2US] = 0, /* XXX */
> + [TGSI_OPCODE_PK4B] = 0, /* XXX */
> + [TGSI_OPCODE_PK4UB] = 0, /* XXX */
> + [TGSI_OPCODE_SEQ] = nir_op_seq,
> + [TGSI_OPCODE_SGT] = 0,
> + [TGSI_OPCODE_SIN] = nir_op_fsin,
> + [TGSI_OPCODE_SLE] = 0,
> + [TGSI_OPCODE_TEX] = 0,
> + [TGSI_OPCODE_TXD] = 0,
> + [TGSI_OPCODE_TXP] = 0,
> + [TGSI_OPCODE_UP2H] = 0, /* XXX */
> + [TGSI_OPCODE_UP2US] = 0, /* XXX */
> + [TGSI_OPCODE_UP4B] = 0, /* XXX */
> + [TGSI_OPCODE_UP4UB] = 0, /* XXX */
> + [TGSI_OPCODE_ARR] = 0,
> +
> + /* No function calls, yet. */
> + [TGSI_OPCODE_CAL] = 0, /* XXX */
> + [TGSI_OPCODE_RET] = 0, /* XXX */
> +
> + [TGSI_OPCODE_SSG] = nir_op_fsign,
> + [TGSI_OPCODE_CMP] = 0,
> + [TGSI_OPCODE_SCS] = 0,
> + [TGSI_OPCODE_TXB] = 0,
> + [TGSI_OPCODE_DIV] = nir_op_fdiv,
> + [TGSI_OPCODE_DP2] = 0,
> + [TGSI_OPCODE_DP2A] = 0,
> + [TGSI_OPCODE_TXL] = 0,
> +
> + [TGSI_OPCODE_BRK] = 0,
> + [TGSI_OPCODE_IF] = 0,
> + [TGSI_OPCODE_UIF] = 0,
> + [TGSI_OPCODE_ELSE] = 0,
> + [TGSI_OPCODE_ENDIF] = 0,
> +
> + [TGSI_OPCODE_DDX_FINE] = nir_op_fddx_fine,
> + [TGSI_OPCODE_DDY_FINE] = nir_op_fddy_fine,
> +
> + [TGSI_OPCODE_PUSHA] = 0, /* XXX */
> + [TGSI_OPCODE_POPA] = 0, /* XXX */
> +
> + [TGSI_OPCODE_CEIL] = nir_op_fceil,
> + [TGSI_OPCODE_I2F] = nir_op_i2f,
> + [TGSI_OPCODE_NOT] = nir_op_inot,
> + [TGSI_OPCODE_TRUNC] = nir_op_ftrunc,
> + [TGSI_OPCODE_SHL] = nir_op_ishl,
> + [TGSI_OPCODE_AND] = nir_op_iand,
> + [TGSI_OPCODE_OR] = nir_op_ior,
> + [TGSI_OPCODE_MOD] = nir_op_fmod,
> + [TGSI_OPCODE_XOR] = nir_op_ixor,
> + [TGSI_OPCODE_SAD] = 0, /* XXX */
> + [TGSI_OPCODE_TXF] = 0,
> + [TGSI_OPCODE_TXQ] = 0,
> +
> + [TGSI_OPCODE_CONT] = 0,
> +
> + [TGSI_OPCODE_EMIT] = 0, /* XXX */
> + [TGSI_OPCODE_ENDPRIM] = 0, /* XXX */
> +
> + [TGSI_OPCODE_BGNLOOP] = 0,
> + [TGSI_OPCODE_BGNSUB] = 0, /* XXX: no function calls */
> + [TGSI_OPCODE_ENDLOOP] = 0,
> + [TGSI_OPCODE_ENDSUB] = 0, /* XXX: no function calls */
> +
> + [TGSI_OPCODE_TXQ_LZ] = 0,
> + [TGSI_OPCODE_NOP] = 0,
> + [TGSI_OPCODE_FSEQ] = nir_op_feq,
> + [TGSI_OPCODE_FSGE] = nir_op_fge,
> + [TGSI_OPCODE_FSLT] = nir_op_flt,
> + [TGSI_OPCODE_FSNE] = nir_op_fne,
> +
> + /* No control flow yet */
> + [TGSI_OPCODE_CALLNZ] = 0, /* XXX */
> + [TGSI_OPCODE_BREAKC] = 0, /* not emitted by glsl_to_tgsi.cpp */
> +
> + [TGSI_OPCODE_KILL_IF] = 0,
> +
> + [TGSI_OPCODE_END] = 0,
> +
> + [TGSI_OPCODE_F2I] = nir_op_f2i,
> + [TGSI_OPCODE_IDIV] = nir_op_idiv,
> + [TGSI_OPCODE_IMAX] = nir_op_imax,
> + [TGSI_OPCODE_IMIN] = nir_op_imin,
> + [TGSI_OPCODE_INEG] = nir_op_ineg,
> + [TGSI_OPCODE_ISGE] = nir_op_ige,
> + [TGSI_OPCODE_ISHR] = nir_op_ishr,
> + [TGSI_OPCODE_ISLT] = nir_op_ilt,
> + [TGSI_OPCODE_F2U] = nir_op_f2u,
> + [TGSI_OPCODE_U2F] = nir_op_u2f,
> + [TGSI_OPCODE_UADD] = nir_op_iadd,
> + [TGSI_OPCODE_UDIV] = nir_op_udiv,
> + [TGSI_OPCODE_UMAD] = 0, /* XXX */
> + [TGSI_OPCODE_UMAX] = nir_op_umax,
> + [TGSI_OPCODE_UMIN] = nir_op_umin,
> + [TGSI_OPCODE_UMOD] = nir_op_umod,
> + [TGSI_OPCODE_UMUL] = nir_op_imul,
> + [TGSI_OPCODE_USEQ] = nir_op_ieq,
> + [TGSI_OPCODE_USGE] = nir_op_uge,
> + [TGSI_OPCODE_USHR] = nir_op_ushr,
> + [TGSI_OPCODE_USLT] = nir_op_ult,
> + [TGSI_OPCODE_USNE] = nir_op_ine,
> +
> + [TGSI_OPCODE_SWITCH] = 0, /* not emitted by glsl_to_tgsi.cpp */
> + [TGSI_OPCODE_CASE] = 0, /* not emitted by glsl_to_tgsi.cpp */
> + [TGSI_OPCODE_DEFAULT] = 0, /* not emitted by glsl_to_tgsi.cpp */
> + [TGSI_OPCODE_ENDSWITCH] = 0, /* not emitted by glsl_to_tgsi.cpp */
> +
> + /* XXX: SAMPLE opcodes */
> +
> + [TGSI_OPCODE_UARL] = nir_op_imov,
> + [TGSI_OPCODE_UCMP] = 0,
> + [TGSI_OPCODE_IABS] = nir_op_iabs,
> + [TGSI_OPCODE_ISSG] = nir_op_isign,
> +
> + /* XXX: atomics */
> +
> + [TGSI_OPCODE_TEX2] = 0,
> + [TGSI_OPCODE_TXB2] = 0,
> + [TGSI_OPCODE_TXL2] = 0,
> +
> + [TGSI_OPCODE_IMUL_HI] = nir_op_imul_high,
> + [TGSI_OPCODE_UMUL_HI] = nir_op_umul_high,
> +
> + [TGSI_OPCODE_TG4] = 0,
> + [TGSI_OPCODE_LODQ] = 0, /* XXX */
> +
> + [TGSI_OPCODE_IBFE] = nir_op_ibitfield_extract,
> + [TGSI_OPCODE_UBFE] = nir_op_ubitfield_extract,
> + [TGSI_OPCODE_BFI] = nir_op_bitfield_insert,
> + [TGSI_OPCODE_BREV] = nir_op_bitfield_reverse,
> + [TGSI_OPCODE_POPC] = nir_op_bit_count,
> + [TGSI_OPCODE_LSB] = nir_op_find_lsb,
> + [TGSI_OPCODE_IMSB] = nir_op_ifind_msb,
> + [TGSI_OPCODE_UMSB] = nir_op_ifind_msb, /* XXX: signed vs unsigned */
> +
> + [TGSI_OPCODE_INTERP_CENTROID] = 0, /* XXX */
> + [TGSI_OPCODE_INTERP_SAMPLE] = 0, /* XXX */
> + [TGSI_OPCODE_INTERP_OFFSET] = 0, /* XXX */
> +};
> +
> +static void
> +ttn_emit_instruction(struct ttn_compile *c)
> +{
> + nir_builder *b = &c->build;
> + struct tgsi_full_instruction *tgsi_inst = &c->token->FullInstruction;
> + unsigned i;
> + unsigned tgsi_op = tgsi_inst->Instruction.Opcode;
> +
> + if (tgsi_op == TGSI_OPCODE_END)
> + return;
> +
> + nir_ssa_def *src[TGSI_FULL_MAX_SRC_REGISTERS];
> + for (i = 0; i < TGSI_FULL_MAX_SRC_REGISTERS; i++) {
> + src[i] = ttn_get_src(c, &tgsi_inst->Src[i]);
> + }
> + nir_alu_dest dest = ttn_get_dest(c, &tgsi_inst->Dst[0]);
> +
> + switch (tgsi_op) {
> + case TGSI_OPCODE_RSQ:
> + ttn_move_dest(b, dest, nir_frsq(b, ttn_channel(b, src[0], X)));
> + break;
> +
> + case TGSI_OPCODE_SQRT:
> + ttn_move_dest(b, dest, nir_fsqrt(b, ttn_channel(b, src[0], X)));
> + break;
> +
> + case TGSI_OPCODE_RCP:
> + ttn_move_dest(b, dest, nir_frcp(b, ttn_channel(b, src[0], X)));
> + break;
> +
> + case TGSI_OPCODE_EX2:
> + ttn_move_dest(b, dest, nir_fexp2(b, ttn_channel(b, src[0], X)));
> + break;
> +
> + case TGSI_OPCODE_LG2:
> + ttn_move_dest(b, dest, nir_flog2(b, ttn_channel(b, src[0], X)));
> + break;
> +
> + case TGSI_OPCODE_POW:
> + ttn_move_dest(b, dest, nir_fpow(b,
> + ttn_channel(b, src[0], X),
> + ttn_channel(b, src[1], X)));
> + break;
> +
> + case TGSI_OPCODE_COS:
> + ttn_move_dest(b, dest, nir_fcos(b, ttn_channel(b, src[0], X)));
> + break;
> +
> + case TGSI_OPCODE_SIN:
> + ttn_move_dest(b, dest, nir_fsin(b, ttn_channel(b, src[0], X)));
> + break;
> +
> + case TGSI_OPCODE_ARL:
> + ttn_arl(b, op_trans[tgsi_op], dest, src);
> + break;
> +
> + case TGSI_OPCODE_EXP:
> + ttn_exp(b, op_trans[tgsi_op], dest, src);
> + break;
> +
> + case TGSI_OPCODE_LOG:
> + ttn_log(b, op_trans[tgsi_op], dest, src);
> + break;
> +
> + case TGSI_OPCODE_DST:
> + ttn_dst(b, op_trans[tgsi_op], dest, src);
> + break;
> +
> + case TGSI_OPCODE_LIT:
> + ttn_lit(b, op_trans[tgsi_op], dest, src);
> + break;
> +
> + case TGSI_OPCODE_CLAMP:
> + ttn_clamp(b, op_trans[tgsi_op], dest, src);
> + break;
> +
> + case TGSI_OPCODE_XPD:
> + ttn_xpd(b, op_trans[tgsi_op], dest, src);
> + break;
> +
> + case TGSI_OPCODE_DP2:
> + ttn_dp2(b, op_trans[tgsi_op], dest, src);
> + break;
> +
> + case TGSI_OPCODE_DP3:
> + ttn_dp3(b, op_trans[tgsi_op], dest, src);
> + break;
> +
> + case TGSI_OPCODE_DP4:
> + ttn_dp4(b, op_trans[tgsi_op], dest, src);
> + break;
> +
> + case TGSI_OPCODE_DP2A:
> + ttn_dp2a(b, op_trans[tgsi_op], dest, src);
> + break;
> +
> + case TGSI_OPCODE_DPH:
> + ttn_dph(b, op_trans[tgsi_op], dest, src);
> + break;
> +
> + case TGSI_OPCODE_LRP:
> + ttn_move_dest(b, dest, nir_flrp(b, src[2], src[1], src[0]));
> + break;
> +
> + case TGSI_OPCODE_KILL:
> + ttn_kill(b, op_trans[tgsi_op], dest, src);
> + break;
> +
> + case TGSI_OPCODE_ARR:
> + ttn_arr(b, op_trans[tgsi_op], dest, src);
> + break;
> +
> + case TGSI_OPCODE_CMP:
> + ttn_cmp(b, op_trans[tgsi_op], dest, src);
> + break;
> +
> + case TGSI_OPCODE_UCMP:
> + ttn_ucmp(b, op_trans[tgsi_op], dest, src);
> + break;
> +
> + case TGSI_OPCODE_SCS:
> + ttn_scs(b, op_trans[tgsi_op], dest, src);
> + break;
> +
> + case TGSI_OPCODE_SGT:
> + ttn_sgt(b, op_trans[tgsi_op], dest, src);
> + break;
> +
> + case TGSI_OPCODE_SLE:
> + ttn_sle(b, op_trans[tgsi_op], dest, src);
> + break;
> +
> + case TGSI_OPCODE_KILL_IF:
> + ttn_kill_if(b, op_trans[tgsi_op], dest, src);
> + break;
> +
> + case TGSI_OPCODE_TEX:
> + case TGSI_OPCODE_TXP:
> + case TGSI_OPCODE_TXL:
> + case TGSI_OPCODE_TXB:
> + case TGSI_OPCODE_TXD:
> + case TGSI_OPCODE_TXQ:
> + case TGSI_OPCODE_TXL2:
> + case TGSI_OPCODE_TXB2:
> + case TGSI_OPCODE_TXQ_LZ:
> + case TGSI_OPCODE_TXF:
> + case TGSI_OPCODE_TG4:
> + ttn_tex(c, dest, src);
> + break;
> +
> + case TGSI_OPCODE_NOP:
> + break;
> +
> + case TGSI_OPCODE_IF:
> + ttn_if(c, src[0], false);
> + break;
> +
> + case TGSI_OPCODE_UIF:
> + ttn_if(c, src[0], true);
> + break;
> +
> + case TGSI_OPCODE_ELSE:
> + ttn_else(c);
> + break;
> +
> + case TGSI_OPCODE_ENDIF:
> + ttn_endif(c);
> + break;
> +
> + case TGSI_OPCODE_BGNLOOP:
> + ttn_bgnloop(c);
> + break;
> +
> + case TGSI_OPCODE_BRK:
> + ttn_brk(b);
> + break;
> +
> + case TGSI_OPCODE_CONT:
> + ttn_cont(b);
> + break;
> +
> + case TGSI_OPCODE_ENDLOOP:
> + ttn_endloop(c);
> + break;
> +
> + default:
> + if (op_trans[tgsi_op] != 0 || tgsi_op == TGSI_OPCODE_MOV) {
> + ttn_alu(b, op_trans[tgsi_op], dest, src);
> + } else {
> + fprintf(stderr, "unknown TGSI opcode: %s\n",
> + tgsi_get_opcode_name(tgsi_op));
> + abort();
> + }
> + break;
> + }
> +
> + if (tgsi_inst->Instruction.Saturate) {
> + assert(tgsi_inst->Instruction.Saturate == TGSI_SAT_ZERO_ONE);
> + assert(!dest.dest.is_ssa);
> + ttn_move_dest(b, dest, nir_fsat(b, ttn_src_for_dest(b, &dest)));
> + }
> +}
> +
> +/**
> + * Puts a NIR intrinsic to store of each TGSI_FILE_OUTPUT value to the
output
> + * variables at the end of the shader.
> + *
> + * We don't generate these incrementally as the TGSI_FILE_OUTPUT values are
> + * written, because there's no output load intrinsic, which means we
couldn't
> + * handle writemasks.
> + */
> +static void
> +ttn_add_output_stores(struct ttn_compile *c)
> +{
> + nir_builder *b = &c->build;
> +
> + foreach_list_typed(nir_variable, var, node, &b->shader->outputs) {
> + unsigned array_len = MAX2(glsl_get_length(var->type), 1);
> + unsigned i;
> +
> + for (i = 0; i < array_len; i++) {
> + nir_intrinsic_instr *store =
> + nir_intrinsic_instr_create(b->shader,
nir_intrinsic_store_output);
> + store->num_components = 4;
> + store->const_index[0] = var->data.driver_location + i;
> + store->const_index[1] = 1;
> + store->src[0].reg.reg = c->output_regs[var-
>data.driver_location].reg;
> + nir_instr_insert_after_cf_list(b->cf_node_list, &store->instr);
> + }
> + }
> +}
> +
> +struct nir_shader *
> +tgsi_to_nir(const void *tgsi_tokens,
> + const nir_shader_compiler_options *options)
> +{
> + struct tgsi_parse_context parser;
> + struct tgsi_shader_info scan;
> + struct ttn_compile *c;
> + struct nir_shader *s;
> + int ret;
> +
> + c = rzalloc(NULL, struct ttn_compile);
> + s = nir_shader_create(NULL, options);
> +
> + nir_function *func = nir_function_create(s, "main");
> + nir_function_overload *overload = nir_function_overload_create(func);
> + nir_function_impl *impl = nir_function_impl_create(overload);
> +
> + nir_builder_init(&c->build, impl);
> + nir_builder_insert_after_cf_list(&c->build, &impl->body);
> +
> + tgsi_scan_shader(tgsi_tokens, &scan);
> + c->scan = &scan;
> +
> + c->output_regs = rzalloc_array(c, struct ttn_reg_info,
> + scan.file_max[TGSI_FILE_OUTPUT] + 1);
> + c->temp_regs = rzalloc_array(c, struct ttn_reg_info,
> + scan.file_max[TGSI_FILE_TEMPORARY] + 1);
> + c->imm_defs = rzalloc_array(c, nir_ssa_def *,
> + scan.file_max[TGSI_FILE_IMMEDIATE] + 1);
> +
> + c->if_stack = rzalloc_array(c, struct exec_list *,
> + (scan.opcode_count[TGSI_OPCODE_IF] +
> + scan.opcode_count[TGSI_OPCODE_UIF]) * 2);
> + c->loop_stack = rzalloc_array(c, struct exec_list *,
> + scan.opcode_count[TGSI_OPCODE_BGNLOOP]);
> +
> + ret = tgsi_parse_init(&parser, tgsi_tokens);
> + assert(ret == TGSI_PARSE_OK);
> +
> + while (!tgsi_parse_end_of_tokens(&parser)) {
> + tgsi_parse_token(&parser);
> + c->token = &parser.FullToken;
> +
> + switch (parser.FullToken.Token.Type) {
> + case TGSI_TOKEN_TYPE_DECLARATION:
> + ttn_emit_declaration(c);
> + break;
> +
> + case TGSI_TOKEN_TYPE_INSTRUCTION:
> + ttn_emit_instruction(c);
> + break;
> +
> + case TGSI_TOKEN_TYPE_IMMEDIATE:
> + ttn_emit_immediate(c);
> + break;
> + }
> + }
> +
> + tgsi_parse_free(&parser);
> +
> + ttn_add_output_stores(c);
> +
> + ralloc_free(c);
> + return s;
> +}
> diff --git a/src/gallium/auxiliary/nir/tgsi_to_nir.h
b/src/gallium/auxiliary/nir/tgsi_to_nir.h
> new file mode 100644
> index 0000000..687348a
> --- /dev/null
> +++ b/src/gallium/auxiliary/nir/tgsi_to_nir.h
> @@ -0,0 +1,30 @@
> +/*
> + * Copyright © 2014 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 "glsl/nir/nir.h"
> +
> +struct nir_shader_compiler_options *options;
> +
> +struct nir_shader *
> +tgsi_to_nir(const void *tgsi_tokens,
> + const struct nir_shader_compiler_options *options);
>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 819 bytes
Desc: This is a digitally signed message part.
URL: <http://lists.freedesktop.org/archives/mesa-dev/attachments/20150327/421a2b54/attachment-0001.sig>
More information about the mesa-dev
mailing list