*/
#include <inttypes.h>
-#include "pipe/p_state.h"
#include "util/u_format.h"
-#include "util/u_hash_table.h"
#include "util/u_hash.h"
+#include "util/u_math.h"
#include "util/u_memory.h"
-#include "util/u_pack_color.h"
-#include "util/format_srgb.h"
#include "util/ralloc.h"
#include "util/hash_table.h"
#include "tgsi/tgsi_dump.h"
#include "tgsi/tgsi_info.h"
#include "tgsi/tgsi_lowering.h"
+#include "tgsi/tgsi_parse.h"
+#include "nir/tgsi_to_nir.h"
#include "vc4_context.h"
#include "vc4_qpu.h"
#include "simpenrose/simpenrose.h"
#endif
-struct vc4_key {
- struct vc4_uncompiled_shader *shader_state;
- struct {
- enum pipe_format format;
- unsigned compare_mode:1;
- unsigned compare_func:3;
- unsigned wrap_s:3;
- unsigned wrap_t:3;
- uint8_t swizzle[4];
- } tex[VC4_MAX_TEXTURE_SAMPLERS];
- uint8_t ucp_enables;
-};
-
-struct vc4_fs_key {
- struct vc4_key base;
- enum pipe_format color_format;
- bool depth_enabled;
- bool stencil_enabled;
- bool stencil_twoside;
- bool stencil_full_writemasks;
- bool is_points;
- bool is_lines;
- bool alpha_test;
- bool point_coord_upper_left;
- bool light_twoside;
- uint8_t alpha_test_func;
- uint32_t point_sprite_mask;
-
- struct pipe_rt_blend_state blend;
-};
-
-struct vc4_vs_key {
- struct vc4_key base;
-
- /**
- * This is a proxy for the array of FS input semantics, which is
- * larger than we would want to put in the key.
- */
- uint64_t compiled_fs_id;
-
- enum pipe_format attr_formats[8];
- bool is_coord;
- bool per_vertex_point_size;
-};
+static struct qreg
+ntq_get_src(struct vc4_compile *c, nir_src src, int i);
static void
resize_qreg_array(struct vc4_compile *c,
}
static struct qreg
-add_uniform(struct vc4_compile *c,
- enum quniform_contents contents,
- uint32_t data)
-{
- uint32_t uniform = c->num_uniforms++;
- struct qreg u = { QFILE_UNIF, uniform };
+indirect_uniform_load(struct vc4_compile *c, nir_intrinsic_instr *intr)
+{
+ struct qreg indirect_offset = ntq_get_src(c, intr->src[0], 0);
+ uint32_t offset = intr->const_index[0];
+ struct vc4_compiler_ubo_range *range = NULL;
+ unsigned i;
+ for (i = 0; i < c->num_uniform_ranges; i++) {
+ range = &c->ubo_ranges[i];
+ if (offset >= range->src_offset &&
+ offset < range->src_offset + range->size) {
+ break;
+ }
+ }
+ /* The driver-location-based offset always has to be within a declared
+ * uniform range.
+ */
+ assert(range);
+ if (!range->used) {
+ range->used = true;
+ range->dst_offset = c->next_ubo_dst_offset;
+ c->next_ubo_dst_offset += range->size;
+ c->num_ubo_ranges++;
+ };
- if (uniform >= c->uniform_array_size) {
- c->uniform_array_size = MAX2(MAX2(16, uniform + 1),
- c->uniform_array_size * 2);
+ offset -= range->src_offset;
- c->uniform_data = reralloc(c, c->uniform_data,
- uint32_t,
- c->uniform_array_size);
- c->uniform_contents = reralloc(c, c->uniform_contents,
- enum quniform_contents,
- c->uniform_array_size);
- }
+ /* Adjust for where we stored the TGSI register base. */
+ indirect_offset = qir_ADD(c, indirect_offset,
+ qir_uniform_ui(c, (range->dst_offset +
+ offset)));
- c->uniform_contents[uniform] = contents;
- c->uniform_data[uniform] = data;
+ /* Clamp to [0, array size). Note that MIN/MAX are signed. */
+ indirect_offset = qir_MAX(c, indirect_offset, qir_uniform_ui(c, 0));
+ indirect_offset = qir_MIN(c, indirect_offset,
+ qir_uniform_ui(c, (range->dst_offset +
+ range->size - 4)));
- return u;
+ qir_TEX_DIRECT(c, indirect_offset, qir_uniform(c, QUNIFORM_UBO_ADDR, 0));
+ c->num_texture_samples++;
+ return qir_TEX_RESULT(c);
}
-static struct qreg
-get_temp_for_uniform(struct vc4_compile *c, enum quniform_contents contents,
- uint32_t data)
+static struct qreg *
+ntq_init_ssa_def(struct vc4_compile *c, nir_ssa_def *def)
{
- struct qreg u = add_uniform(c, contents, data);
- struct qreg t = qir_MOV(c, u);
- return t;
+ struct qreg *qregs = ralloc_array(c->def_ht, struct qreg,
+ def->num_components);
+ _mesa_hash_table_insert(c->def_ht, def, qregs);
+ return qregs;
}
-static struct qreg
-qir_uniform_ui(struct vc4_compile *c, uint32_t ui)
+static struct qreg *
+ntq_get_dest(struct vc4_compile *c, nir_dest *dest)
{
- return get_temp_for_uniform(c, QUNIFORM_CONSTANT, ui);
+ if (dest->is_ssa) {
+ struct qreg *qregs = ntq_init_ssa_def(c, &dest->ssa);
+ for (int i = 0; i < dest->ssa.num_components; i++)
+ qregs[i] = c->undef;
+ return qregs;
+ } else {
+ nir_register *reg = dest->reg.reg;
+ assert(dest->reg.base_offset == 0);
+ assert(reg->num_array_elems == 0);
+ struct hash_entry *entry =
+ _mesa_hash_table_search(c->def_ht, reg);
+ return entry->data;
+ }
}
static struct qreg
-qir_uniform_f(struct vc4_compile *c, float f)
+ntq_get_src(struct vc4_compile *c, nir_src src, int i)
{
- return qir_uniform_ui(c, fui(f));
+ struct hash_entry *entry;
+ if (src.is_ssa) {
+ entry = _mesa_hash_table_search(c->def_ht, src.ssa);
+ assert(i < src.ssa->num_components);
+ } else {
+ nir_register *reg = src.reg.reg;
+ entry = _mesa_hash_table_search(c->def_ht, reg);
+ assert(reg->num_array_elems == 0);
+ assert(src.reg.base_offset == 0);
+ assert(i < reg->num_components);
+ }
+
+ struct qreg *qregs = entry->data;
+ return qregs[i];
}
static struct qreg
-get_src(struct vc4_compile *c, unsigned tgsi_op,
- struct tgsi_src_register *src, int i)
+ntq_get_alu_src(struct vc4_compile *c, nir_alu_instr *instr,
+ unsigned src)
{
- struct qreg r = c->undef;
-
- uint32_t s = i;
- switch (i) {
- case TGSI_SWIZZLE_X:
- s = src->SwizzleX;
- break;
- case TGSI_SWIZZLE_Y:
- s = src->SwizzleY;
- break;
- case TGSI_SWIZZLE_Z:
- s = src->SwizzleZ;
- break;
- case TGSI_SWIZZLE_W:
- s = src->SwizzleW;
- break;
- default:
- abort();
- }
+ assert(util_is_power_of_two(instr->dest.write_mask));
+ unsigned chan = ffs(instr->dest.write_mask) - 1;
+ struct qreg r = ntq_get_src(c, instr->src[src].src,
+ instr->src[src].swizzle[chan]);
- assert(!src->Indirect);
-
- switch (src->File) {
- case TGSI_FILE_NULL:
- return r;
- case TGSI_FILE_TEMPORARY:
- r = c->temps[src->Index * 4 + s];
- break;
- case TGSI_FILE_IMMEDIATE:
- r = c->consts[src->Index * 4 + s];
- break;
- case TGSI_FILE_CONSTANT:
- r = get_temp_for_uniform(c, QUNIFORM_UNIFORM,
- src->Index * 4 + s);
- break;
- case TGSI_FILE_INPUT:
- r = c->inputs[src->Index * 4 + s];
- break;
- case TGSI_FILE_SAMPLER:
- case TGSI_FILE_SAMPLER_VIEW:
- r = c->undef;
- break;
- default:
- fprintf(stderr, "unknown src file %d\n", src->File);
- abort();
- }
-
- if (src->Absolute)
- r = qir_FMAXABS(c, r, r);
-
- if (src->Negate) {
- switch (tgsi_opcode_infer_src_type(tgsi_op)) {
- case TGSI_TYPE_SIGNED:
- case TGSI_TYPE_UNSIGNED:
- r = qir_SUB(c, qir_uniform_ui(c, 0), r);
- break;
- default:
- r = qir_FSUB(c, qir_uniform_f(c, 0.0), r);
- break;
- }
- }
+ assert(!instr->src[src].abs);
+ assert(!instr->src[src].negate);
return r;
};
-
-static void
-update_dst(struct vc4_compile *c, struct tgsi_full_instruction *tgsi_inst,
- int i, struct qreg val)
-{
- struct tgsi_dst_register *tgsi_dst = &tgsi_inst->Dst[0].Register;
-
- assert(!tgsi_dst->Indirect);
-
- switch (tgsi_dst->File) {
- case TGSI_FILE_TEMPORARY:
- c->temps[tgsi_dst->Index * 4 + i] = val;
- break;
- case TGSI_FILE_OUTPUT:
- c->outputs[tgsi_dst->Index * 4 + i] = val;
- c->num_outputs = MAX2(c->num_outputs,
- tgsi_dst->Index * 4 + i + 1);
- break;
- default:
- fprintf(stderr, "unknown dst file %d\n", tgsi_dst->File);
- abort();
- }
-};
-
static struct qreg
get_swizzled_channel(struct vc4_compile *c,
struct qreg *srcs, int swiz)
}
}
-static struct qreg
-tgsi_to_qir_alu(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst,
- enum qop op, struct qreg *src, int i)
+static inline struct qreg
+qir_SAT(struct vc4_compile *c, struct qreg val)
{
- struct qreg dst = qir_get_temp(c);
- qir_emit(c, qir_inst4(op, dst,
- src[0 * 4 + i],
- src[1 * 4 + i],
- src[2 * 4 + i],
- c->undef));
- return dst;
+ return qir_FMAX(c,
+ qir_FMIN(c, val, qir_uniform_f(c, 1.0)),
+ qir_uniform_f(c, 0.0));
}
static struct qreg
-tgsi_to_qir_scalar(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst,
- enum qop op, struct qreg *src, int i)
+ntq_rcp(struct vc4_compile *c, struct qreg x)
{
- struct qreg dst = qir_get_temp(c);
- qir_emit(c, qir_inst(op, dst,
- src[0 * 4 + 0],
- c->undef));
- return dst;
+ struct qreg r = qir_RCP(c, x);
+
+ /* Apply a Newton-Raphson step to improve the accuracy. */
+ r = qir_FMUL(c, r, qir_FSUB(c,
+ qir_uniform_f(c, 2.0),
+ qir_FMUL(c, x, r)));
+
+ return r;
}
static struct qreg
-tgsi_to_qir_rcp(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst,
- enum qop op, struct qreg *src, int i)
+ntq_rsq(struct vc4_compile *c, struct qreg x)
{
- struct qreg x = src[0 * 4 + 0];
- struct qreg r = qir_RCP(c, x);
+ struct qreg r = qir_RSQ(c, x);
/* Apply a Newton-Raphson step to improve the accuracy. */
r = qir_FMUL(c, r, qir_FSUB(c,
- qir_uniform_f(c, 2.0),
- qir_FMUL(c, x, r)));
+ qir_uniform_f(c, 1.5),
+ qir_FMUL(c,
+ qir_uniform_f(c, 0.5),
+ qir_FMUL(c, x,
+ qir_FMUL(c, r, r)))));
return r;
}
}
static struct qreg
-tgsi_to_qir_umul(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst,
- enum qop op, struct qreg *src, int i)
-{
- struct qreg src0_hi = qir_SHR(c, src[0 * 4 + i],
- qir_uniform_ui(c, 16));
- struct qreg src0_lo = qir_AND(c, src[0 * 4 + i],
- qir_uniform_ui(c, 0xffff));
- struct qreg src1_hi = qir_SHR(c, src[1 * 4 + i],
- qir_uniform_ui(c, 16));
- struct qreg src1_lo = qir_AND(c, src[1 * 4 + i],
- qir_uniform_ui(c, 0xffff));
-
- struct qreg hilo = qir_MUL24(c, src0_hi, src1_lo);
- struct qreg lohi = qir_MUL24(c, src0_lo, src1_hi);
- struct qreg lolo = qir_MUL24(c, src0_lo, src1_lo);
-
- return qir_ADD(c, lolo, qir_SHL(c,
- qir_ADD(c, hilo, lohi),
- qir_uniform_ui(c, 16)));
-}
-
-static struct qreg
-tgsi_to_qir_idiv(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst,
- enum qop op, struct qreg *src, int i)
-{
- return qir_FTOI(c, qir_FMUL(c,
- qir_ITOF(c, src[0 * 4 + i]),
- qir_RCP(c, qir_ITOF(c, src[1 * 4 + i]))));
-}
-
-static struct qreg
-tgsi_to_qir_ineg(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst,
- enum qop op, struct qreg *src, int i)
-{
- return qir_SUB(c, qir_uniform_ui(c, 0), src[0 * 4 + i]);
-}
-
-static struct qreg
-tgsi_to_qir_seq(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst,
- enum qop op, struct qreg *src, int i)
-{
- qir_SF(c, qir_FSUB(c, src[0 * 4 + i], src[1 * 4 + i]));
- return qir_SEL_X_0_ZS(c, qir_uniform_f(c, 1.0));
-}
-
-static struct qreg
-tgsi_to_qir_sne(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst,
- enum qop op, struct qreg *src, int i)
-{
- qir_SF(c, qir_FSUB(c, src[0 * 4 + i], src[1 * 4 + i]));
- return qir_SEL_X_0_ZC(c, qir_uniform_f(c, 1.0));
-}
-
-static struct qreg
-tgsi_to_qir_slt(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst,
- enum qop op, struct qreg *src, int i)
-{
- qir_SF(c, qir_FSUB(c, src[0 * 4 + i], src[1 * 4 + i]));
- return qir_SEL_X_0_NS(c, qir_uniform_f(c, 1.0));
-}
-
-static struct qreg
-tgsi_to_qir_sge(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst,
- enum qop op, struct qreg *src, int i)
-{
- qir_SF(c, qir_FSUB(c, src[0 * 4 + i], src[1 * 4 + i]));
- return qir_SEL_X_0_NC(c, qir_uniform_f(c, 1.0));
-}
-
-static struct qreg
-tgsi_to_qir_fseq(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst,
- enum qop op, struct qreg *src, int i)
-{
- qir_SF(c, qir_FSUB(c, src[0 * 4 + i], src[1 * 4 + i]));
- return qir_SEL_X_0_ZS(c, qir_uniform_ui(c, ~0));
-}
-
-static struct qreg
-tgsi_to_qir_fsne(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst,
- enum qop op, struct qreg *src, int i)
-{
- qir_SF(c, qir_FSUB(c, src[0 * 4 + i], src[1 * 4 + i]));
- return qir_SEL_X_0_ZC(c, qir_uniform_ui(c, ~0));
-}
-
-static struct qreg
-tgsi_to_qir_fslt(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst,
- enum qop op, struct qreg *src, int i)
-{
- qir_SF(c, qir_FSUB(c, src[0 * 4 + i], src[1 * 4 + i]));
- return qir_SEL_X_0_NS(c, qir_uniform_ui(c, ~0));
-}
-
-static struct qreg
-tgsi_to_qir_fsge(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst,
- enum qop op, struct qreg *src, int i)
-{
- qir_SF(c, qir_FSUB(c, src[0 * 4 + i], src[1 * 4 + i]));
- return qir_SEL_X_0_NC(c, qir_uniform_ui(c, ~0));
-}
-
-static struct qreg
-tgsi_to_qir_useq(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst,
- enum qop op, struct qreg *src, int i)
-{
- qir_SF(c, qir_SUB(c, src[0 * 4 + i], src[1 * 4 + i]));
- return qir_SEL_X_0_ZS(c, qir_uniform_ui(c, ~0));
-}
-
-static struct qreg
-tgsi_to_qir_usne(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst,
- enum qop op, struct qreg *src, int i)
-{
- qir_SF(c, qir_SUB(c, src[0 * 4 + i], src[1 * 4 + i]));
- return qir_SEL_X_0_ZC(c, qir_uniform_ui(c, ~0));
-}
-
-static struct qreg
-tgsi_to_qir_islt(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst,
- enum qop op, struct qreg *src, int i)
-{
- qir_SF(c, qir_SUB(c, src[0 * 4 + i], src[1 * 4 + i]));
- return qir_SEL_X_0_NS(c, qir_uniform_ui(c, ~0));
-}
-
-static struct qreg
-tgsi_to_qir_isge(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst,
- enum qop op, struct qreg *src, int i)
-{
- qir_SF(c, qir_SUB(c, src[0 * 4 + i], src[1 * 4 + i]));
- return qir_SEL_X_0_NC(c, qir_uniform_ui(c, ~0));
-}
-
-static struct qreg
-tgsi_to_qir_cmp(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst,
- enum qop op, struct qreg *src, int i)
-{
- qir_SF(c, src[0 * 4 + i]);
- return qir_SEL_X_Y_NS(c,
- src[1 * 4 + i],
- src[2 * 4 + i]);
-}
-
-static struct qreg
-tgsi_to_qir_mad(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst,
- enum qop op, struct qreg *src, int i)
+ntq_umul(struct vc4_compile *c, struct qreg src0, struct qreg src1)
{
- return qir_FADD(c,
- qir_FMUL(c,
- src[0 * 4 + i],
- src[1 * 4 + i]),
- src[2 * 4 + i]);
-}
+ struct qreg src0_hi = qir_SHR(c, src0,
+ qir_uniform_ui(c, 24));
+ struct qreg src1_hi = qir_SHR(c, src1,
+ qir_uniform_ui(c, 24));
-static struct qreg
-tgsi_to_qir_lrp(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst,
- enum qop op, struct qreg *src, int i)
-{
- struct qreg src0 = src[0 * 4 + i];
- struct qreg src1 = src[1 * 4 + i];
- struct qreg src2 = src[2 * 4 + i];
-
- /* LRP is:
- * src0 * src1 + (1 - src0) * src2.
- * -> src0 * src1 + src2 - src0 * src2
- * -> src2 + src0 * (src1 - src2)
- */
- return qir_FADD(c, src2, qir_FMUL(c, src0, qir_FSUB(c, src1, src2)));
+ struct qreg hilo = qir_MUL24(c, src0_hi, src1);
+ struct qreg lohi = qir_MUL24(c, src0, src1_hi);
+ struct qreg lolo = qir_MUL24(c, src0, src1);
+ return qir_ADD(c, lolo, qir_SHL(c,
+ qir_ADD(c, hilo, lohi),
+ qir_uniform_ui(c, 24)));
}
static void
-tgsi_to_qir_tex(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst,
- enum qop op, struct qreg *src)
-{
- assert(!tgsi_inst->Instruction.Saturate);
-
- struct qreg s = src[0 * 4 + 0];
- struct qreg t = src[0 * 4 + 1];
- struct qreg r = src[0 * 4 + 2];
- uint32_t unit = tgsi_inst->Src[1].Register.Index;
- bool is_txl = tgsi_inst->Instruction.Opcode == TGSI_OPCODE_TXL;
-
- struct qreg proj = c->undef;
- if (tgsi_inst->Instruction.Opcode == TGSI_OPCODE_TXP) {
- proj = qir_RCP(c, src[0 * 4 + 3]);
- s = qir_FMUL(c, s, proj);
- t = qir_FMUL(c, t, proj);
+ntq_emit_tex(struct vc4_compile *c, nir_tex_instr *instr)
+{
+ struct qreg s, t, r, lod, proj, compare;
+ bool is_txb = false, is_txl = false, has_proj = false;
+ unsigned unit = instr->sampler_index;
+
+ for (unsigned i = 0; i < instr->num_srcs; i++) {
+ switch (instr->src[i].src_type) {
+ case nir_tex_src_coord:
+ s = ntq_get_src(c, instr->src[i].src, 0);
+ if (instr->sampler_dim == GLSL_SAMPLER_DIM_1D)
+ t = qir_uniform_f(c, 0.5);
+ else
+ t = ntq_get_src(c, instr->src[i].src, 1);
+ if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE)
+ r = ntq_get_src(c, instr->src[i].src, 2);
+ break;
+ case nir_tex_src_bias:
+ lod = ntq_get_src(c, instr->src[i].src, 0);
+ is_txb = true;
+ break;
+ case nir_tex_src_lod:
+ lod = ntq_get_src(c, instr->src[i].src, 0);
+ is_txl = true;
+ break;
+ case nir_tex_src_comparitor:
+ compare = ntq_get_src(c, instr->src[i].src, 0);
+ break;
+ case nir_tex_src_projector:
+ proj = qir_RCP(c, ntq_get_src(c, instr->src[i].src, 0));
+ s = qir_FMUL(c, s, proj);
+ t = qir_FMUL(c, t, proj);
+ has_proj = true;
+ break;
+ default:
+ unreachable("unknown texture source");
+ }
}
struct qreg texture_u[] = {
- add_uniform(c, QUNIFORM_TEXTURE_CONFIG_P0, unit),
- add_uniform(c, QUNIFORM_TEXTURE_CONFIG_P1, unit),
- add_uniform(c, QUNIFORM_CONSTANT, 0),
- add_uniform(c, QUNIFORM_CONSTANT, 0),
+ qir_uniform(c, QUNIFORM_TEXTURE_CONFIG_P0, unit),
+ qir_uniform(c, QUNIFORM_TEXTURE_CONFIG_P1, unit),
+ qir_uniform(c, QUNIFORM_CONSTANT, 0),
+ qir_uniform(c, QUNIFORM_CONSTANT, 0),
};
uint32_t next_texture_u = 0;
* we have to rescale from ([0, width], [0, height]) to ([0, 1], [0,
* 1]).
*/
- if (tgsi_inst->Texture.Texture == TGSI_TEXTURE_RECT ||
- tgsi_inst->Texture.Texture == TGSI_TEXTURE_SHADOWRECT) {
+ if (instr->sampler_dim == GLSL_SAMPLER_DIM_RECT) {
s = qir_FMUL(c, s,
- get_temp_for_uniform(c,
- QUNIFORM_TEXRECT_SCALE_X,
- unit));
+ qir_uniform(c, QUNIFORM_TEXRECT_SCALE_X, unit));
t = qir_FMUL(c, t,
- get_temp_for_uniform(c,
- QUNIFORM_TEXRECT_SCALE_Y,
- unit));
+ qir_uniform(c, QUNIFORM_TEXRECT_SCALE_Y, unit));
}
- if (tgsi_inst->Texture.Texture == TGSI_TEXTURE_CUBE ||
- tgsi_inst->Texture.Texture == TGSI_TEXTURE_SHADOWCUBE ||
- is_txl) {
- texture_u[2] = add_uniform(c, QUNIFORM_TEXTURE_CONFIG_P2,
+ if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE || is_txl) {
+ texture_u[2] = qir_uniform(c, QUNIFORM_TEXTURE_CONFIG_P2,
unit | (is_txl << 16));
}
- if (tgsi_inst->Texture.Texture == TGSI_TEXTURE_CUBE ||
- tgsi_inst->Texture.Texture == TGSI_TEXTURE_SHADOWCUBE) {
+ if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE) {
struct qreg ma = qir_FMAXABS(c, qir_FMAXABS(c, s, t), r);
struct qreg rcp_ma = qir_RCP(c, ma);
s = qir_FMUL(c, s, rcp_ma);
c->key->tex[unit].wrap_s == PIPE_TEX_WRAP_CLAMP ||
c->key->tex[unit].wrap_t == PIPE_TEX_WRAP_CLAMP_TO_BORDER ||
c->key->tex[unit].wrap_t == PIPE_TEX_WRAP_CLAMP) {
- qir_TEX_R(c, get_temp_for_uniform(c, QUNIFORM_TEXTURE_BORDER_COLOR, unit),
+ qir_TEX_R(c, qir_uniform(c, QUNIFORM_TEXTURE_BORDER_COLOR, unit),
texture_u[next_texture_u++]);
}
if (c->key->tex[unit].wrap_s == PIPE_TEX_WRAP_CLAMP) {
- s = qir_FMIN(c, qir_FMAX(c, s, qir_uniform_f(c, 0.0)),
- qir_uniform_f(c, 1.0));
+ s = qir_SAT(c, s);
}
if (c->key->tex[unit].wrap_t == PIPE_TEX_WRAP_CLAMP) {
- t = qir_FMIN(c, qir_FMAX(c, t, qir_uniform_f(c, 0.0)),
- qir_uniform_f(c, 1.0));
+ t = qir_SAT(c, t);
}
qir_TEX_T(c, t, texture_u[next_texture_u++]);
- if (tgsi_inst->Instruction.Opcode == TGSI_OPCODE_TXB ||
- tgsi_inst->Instruction.Opcode == TGSI_OPCODE_TXL)
- qir_TEX_B(c, src[0 * 4 + 3], texture_u[next_texture_u++]);
+ if (is_txl || is_txb)
+ qir_TEX_B(c, lod, texture_u[next_texture_u++]);
qir_TEX_S(c, s, texture_u[next_texture_u++]);
c->num_texture_samples++;
- struct qreg r4 = qir_TEX_RESULT(c);
+ struct qreg tex = qir_TEX_RESULT(c);
enum pipe_format format = c->key->tex[unit].format;
struct qreg unpacked[4];
if (util_format_is_depth_or_stencil(format)) {
- struct qreg depthf = qir_ITOF(c, qir_SHR(c, r4,
+ struct qreg depthf = qir_ITOF(c, qir_SHR(c, tex,
qir_uniform_ui(c, 8)));
struct qreg normalized = qir_FMUL(c, depthf,
qir_uniform_f(c, 1.0f/0xffffff));
struct qreg one = qir_uniform_f(c, 1.0f);
if (c->key->tex[unit].compare_mode) {
- struct qreg compare = src[0 * 4 + 2];
-
- if (tgsi_inst->Instruction.Opcode == TGSI_OPCODE_TXP)
+ if (has_proj)
compare = qir_FMUL(c, compare, proj);
switch (c->key->tex[unit].compare_func) {
unpacked[i] = depth_output;
} else {
for (int i = 0; i < 4; i++)
- unpacked[i] = qir_R4_UNPACK(c, r4, i);
+ unpacked[i] = qir_UNPACK_8_F(c, tex, i);
}
const uint8_t *format_swiz = vc4_get_format_swizzle(format);
texture_output[i]);
}
+ struct qreg *dest = ntq_get_dest(c, &instr->dest);
for (int i = 0; i < 4; i++) {
- if (!(tgsi_inst->Dst[0].Register.WriteMask & (1 << i)))
- continue;
-
- update_dst(c, tgsi_inst, i,
- get_swizzled_channel(c, texture_output,
- c->key->tex[unit].swizzle[i]));
+ dest[i] = get_swizzled_channel(c, texture_output,
+ c->key->tex[unit].swizzle[i]);
}
}
-static struct qreg
-tgsi_to_qir_trunc(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst,
- enum qop op, struct qreg *src, int i)
-{
- return qir_ITOF(c, qir_FTOI(c, src[0 * 4 + i]));
-}
-
/**
* Computes x - floor(x), which is tricky because our FTOI truncates (rounds
* to zero).
*/
static struct qreg
-tgsi_to_qir_frc(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst,
- enum qop op, struct qreg *src, int i)
+ntq_ffract(struct vc4_compile *c, struct qreg src)
{
- struct qreg trunc = qir_ITOF(c, qir_FTOI(c, src[0 * 4 + i]));
- struct qreg diff = qir_FSUB(c, src[0 * 4 + i], trunc);
+ struct qreg trunc = qir_ITOF(c, qir_FTOI(c, src));
+ struct qreg diff = qir_FSUB(c, src, trunc);
qir_SF(c, diff);
return qir_SEL_X_Y_NS(c,
qir_FADD(c, diff, qir_uniform_f(c, 1.0)),
* zero).
*/
static struct qreg
-tgsi_to_qir_flr(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst,
- enum qop op, struct qreg *src, int i)
+ntq_ffloor(struct vc4_compile *c, struct qreg src)
{
- struct qreg trunc = qir_ITOF(c, qir_FTOI(c, src[0 * 4 + i]));
+ struct qreg trunc = qir_ITOF(c, qir_FTOI(c, src));
/* This will be < 0 if we truncated and the truncation was of a value
* that was < 0 in the first place.
*/
- qir_SF(c, qir_FSUB(c, src[0 * 4 + i], trunc));
+ qir_SF(c, qir_FSUB(c, src, trunc));
return qir_SEL_X_Y_NS(c,
qir_FSUB(c, trunc, qir_uniform_f(c, 1.0)),
* zero).
*/
static struct qreg
-tgsi_to_qir_ceil(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst,
- enum qop op, struct qreg *src, int i)
+ntq_fceil(struct vc4_compile *c, struct qreg src)
{
- struct qreg trunc = qir_ITOF(c, qir_FTOI(c, src[0 * 4 + i]));
+ struct qreg trunc = qir_ITOF(c, qir_FTOI(c, src));
/* This will be < 0 if we truncated and the truncation was of a value
* that was > 0 in the first place.
*/
- qir_SF(c, qir_FSUB(c, trunc, src[0 * 4 + i]));
+ qir_SF(c, qir_FSUB(c, trunc, src));
return qir_SEL_X_Y_NS(c,
qir_FADD(c, trunc, qir_uniform_f(c, 1.0)),
}
static struct qreg
-tgsi_to_qir_abs(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst,
- enum qop op, struct qreg *src, int i)
-{
- struct qreg arg = src[0 * 4 + i];
- return qir_FMAXABS(c, arg, arg);
-}
-
-/* Note that this instruction replicates its result from the x channel */
-static struct qreg
-tgsi_to_qir_sin(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst,
- enum qop op, struct qreg *src, int i)
+ntq_fsin(struct vc4_compile *c, struct qreg src)
{
float coeff[] = {
-2.0 * M_PI,
struct qreg scaled_x =
qir_FMUL(c,
- src[0 * 4 + 0],
- qir_uniform_f(c, 1.0f / (M_PI * 2.0f)));
+ src,
+ qir_uniform_f(c, 1.0 / (M_PI * 2.0)));
struct qreg x = qir_FADD(c,
- tgsi_to_qir_frc(c, NULL, 0, &scaled_x, 0),
+ ntq_ffract(c, scaled_x),
qir_uniform_f(c, -0.5));
struct qreg x2 = qir_FMUL(c, x, x);
struct qreg sum = qir_FMUL(c, x, qir_uniform_f(c, coeff[0]));
return sum;
}
-/* Note that this instruction replicates its result from the x channel */
static struct qreg
-tgsi_to_qir_cos(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst,
- enum qop op, struct qreg *src, int i)
+ntq_fcos(struct vc4_compile *c, struct qreg src)
{
float coeff[] = {
-1.0f,
};
struct qreg scaled_x =
- qir_FMUL(c, src[0 * 4 + 0],
+ qir_FMUL(c, src,
qir_uniform_f(c, 1.0f / (M_PI * 2.0f)));
struct qreg x_frac = qir_FADD(c,
- tgsi_to_qir_frc(c, NULL, 0, &scaled_x, 0),
+ ntq_ffract(c, scaled_x),
qir_uniform_f(c, -0.5));
struct qreg sum = qir_uniform_f(c, coeff[0]);
}
static struct qreg
-tgsi_to_qir_clamp(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst,
- enum qop op, struct qreg *src, int i)
-{
- return qir_FMAX(c, qir_FMIN(c,
- src[0 * 4 + i],
- src[2 * 4 + i]),
- src[1 * 4 + i]);
-}
-
-static struct qreg
-tgsi_to_qir_ssg(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst,
- enum qop op, struct qreg *src, int i)
+ntq_fsign(struct vc4_compile *c, struct qreg src)
{
- qir_SF(c, src[0 * 4 + i]);
+ qir_SF(c, src);
return qir_SEL_X_Y_NC(c,
qir_SEL_X_0_ZC(c, qir_uniform_f(c, 1.0)),
qir_uniform_f(c, -1.0));
}
+static struct qreg
+get_channel_from_vpm(struct vc4_compile *c,
+ struct qreg *vpm_reads,
+ uint8_t swiz,
+ const struct util_format_description *desc)
+{
+ const struct util_format_channel_description *chan =
+ &desc->channel[swiz];
+ struct qreg temp;
+
+ if (swiz > UTIL_FORMAT_SWIZZLE_W)
+ return get_swizzled_channel(c, vpm_reads, swiz);
+ else if (chan->size == 32 &&
+ chan->type == UTIL_FORMAT_TYPE_FLOAT) {
+ return get_swizzled_channel(c, vpm_reads, swiz);
+ } else if (chan->size == 32 &&
+ chan->type == UTIL_FORMAT_TYPE_SIGNED) {
+ if (chan->normalized) {
+ return qir_FMUL(c,
+ qir_ITOF(c, vpm_reads[swiz]),
+ qir_uniform_f(c,
+ 1.0 / 0x7fffffff));
+ } else {
+ return qir_ITOF(c, vpm_reads[swiz]);
+ }
+ } else if (chan->size == 8 &&
+ (chan->type == UTIL_FORMAT_TYPE_UNSIGNED ||
+ chan->type == UTIL_FORMAT_TYPE_SIGNED)) {
+ struct qreg vpm = vpm_reads[0];
+ if (chan->type == UTIL_FORMAT_TYPE_SIGNED) {
+ temp = qir_XOR(c, vpm, qir_uniform_ui(c, 0x80808080));
+ if (chan->normalized) {
+ return qir_FSUB(c, qir_FMUL(c,
+ qir_UNPACK_8_F(c, temp, swiz),
+ qir_uniform_f(c, 2.0)),
+ qir_uniform_f(c, 1.0));
+ } else {
+ return qir_FADD(c,
+ qir_ITOF(c,
+ qir_UNPACK_8_I(c, temp,
+ swiz)),
+ qir_uniform_f(c, -128.0));
+ }
+ } else {
+ if (chan->normalized) {
+ return qir_UNPACK_8_F(c, vpm, swiz);
+ } else {
+ return qir_ITOF(c, qir_UNPACK_8_I(c, vpm, swiz));
+ }
+ }
+ } else if (chan->size == 16 &&
+ (chan->type == UTIL_FORMAT_TYPE_UNSIGNED ||
+ chan->type == UTIL_FORMAT_TYPE_SIGNED)) {
+ struct qreg vpm = vpm_reads[swiz / 2];
+
+ /* Note that UNPACK_16F eats a half float, not ints, so we use
+ * UNPACK_16_I for all of these.
+ */
+ if (chan->type == UTIL_FORMAT_TYPE_SIGNED) {
+ temp = qir_ITOF(c, qir_UNPACK_16_I(c, vpm, swiz % 2));
+ if (chan->normalized) {
+ return qir_FMUL(c, temp,
+ qir_uniform_f(c, 1/32768.0f));
+ } else {
+ return temp;
+ }
+ } else {
+ /* UNPACK_16I sign-extends, so we have to emit ANDs. */
+ temp = vpm;
+ if (swiz == 1 || swiz == 3)
+ temp = qir_UNPACK_16_I(c, temp, 1);
+ temp = qir_AND(c, temp, qir_uniform_ui(c, 0xffff));
+ temp = qir_ITOF(c, temp);
+
+ if (chan->normalized) {
+ return qir_FMUL(c, temp,
+ qir_uniform_f(c, 1 / 65535.0));
+ } else {
+ return temp;
+ }
+ }
+ } else {
+ return c->undef;
+ }
+}
+
static void
emit_vertex_input(struct vc4_compile *c, int attr)
{
enum pipe_format format = c->vs_key->attr_formats[attr];
+ uint32_t attr_size = util_format_get_blocksize(format);
struct qreg vpm_reads[4];
- /* Right now, we're setting the VPM offsets to be 16 bytes wide every
- * time, so we always read 4 32-bit VPM entries.
- */
- for (int i = 0; i < 4; i++) {
- vpm_reads[i] = qir_get_temp(c);
- qir_emit(c, qir_inst(QOP_VPM_READ,
- vpm_reads[i],
- c->undef,
- c->undef));
+ c->vattr_sizes[attr] = align(attr_size, 4);
+ for (int i = 0; i < align(attr_size, 4) / 4; i++) {
+ struct qreg vpm = { QFILE_VPM, attr * 4 + i };
+ vpm_reads[i] = qir_MOV(c, vpm);
c->num_inputs++;
}
for (int i = 0; i < 4; i++) {
uint8_t swiz = desc->swizzle[i];
- struct qreg result;
-
- if (swiz > UTIL_FORMAT_SWIZZLE_W)
- result = get_swizzled_channel(c, vpm_reads, swiz);
- else if (desc->channel[swiz].size == 32 &&
- desc->channel[swiz].type == UTIL_FORMAT_TYPE_FLOAT) {
- result = get_swizzled_channel(c, vpm_reads, swiz);
- } else if (desc->channel[swiz].size == 8 &&
- (desc->channel[swiz].type == UTIL_FORMAT_TYPE_UNSIGNED ||
- desc->channel[swiz].type == UTIL_FORMAT_TYPE_SIGNED) &&
- desc->channel[swiz].normalized) {
- struct qreg vpm = vpm_reads[0];
- if (desc->channel[swiz].type == UTIL_FORMAT_TYPE_SIGNED)
- vpm = qir_XOR(c, vpm, qir_uniform_ui(c, 0x80808080));
- result = qir_UNPACK_8(c, vpm, swiz);
- } else {
+ struct qreg result = get_channel_from_vpm(c, vpm_reads,
+ swiz, desc);
+
+ if (result.file == QFILE_NULL) {
if (!format_warned) {
fprintf(stderr,
"vtx element %d unsupported type: %s\n",
}
result = qir_uniform_f(c, 0.0);
}
-
- if (desc->channel[swiz].normalized &&
- desc->channel[swiz].type == UTIL_FORMAT_TYPE_SIGNED) {
- result = qir_FSUB(c,
- qir_FMUL(c,
- result,
- qir_uniform_f(c, 2.0)),
- qir_uniform_f(c, 1.0));
- }
-
c->inputs[attr * 4 + i] = result;
}
}
-static void
-tgsi_to_qir_kill_if(struct vc4_compile *c, struct qreg *src, int i)
-{
- if (c->discard.file == QFILE_NULL)
- c->discard = qir_uniform_f(c, 0.0);
- qir_SF(c, src[0 * 4 + i]);
- c->discard = qir_SEL_X_Y_NS(c, qir_uniform_f(c, 1.0),
- c->discard);
-}
-
static void
emit_fragcoord_input(struct vc4_compile *c, int attr)
{
c->inputs[attr * 4 + 3] = qir_RCP(c, qir_FRAG_W(c));
}
-static void
-emit_point_coord_input(struct vc4_compile *c, int attr)
-{
- if (c->point_x.file == QFILE_NULL) {
- c->point_x = qir_uniform_f(c, 0.0);
- c->point_y = qir_uniform_f(c, 0.0);
- }
-
- c->inputs[attr * 4 + 0] = c->point_x;
- if (c->fs_key->point_coord_upper_left) {
- c->inputs[attr * 4 + 1] = qir_FSUB(c,
- qir_uniform_f(c, 1.0),
- c->point_y);
- } else {
- c->inputs[attr * 4 + 1] = c->point_y;
- }
- c->inputs[attr * 4 + 2] = qir_uniform_f(c, 0.0);
- c->inputs[attr * 4 + 3] = qir_uniform_f(c, 1.0);
-}
-
static struct qreg
emit_fragment_varying(struct vc4_compile *c, uint8_t semantic,
uint8_t index, uint8_t swizzle)
static void
emit_fragment_input(struct vc4_compile *c, int attr,
- struct tgsi_full_declaration *decl)
+ unsigned semantic_name, unsigned semantic_index)
{
for (int i = 0; i < 4; i++) {
c->inputs[attr * 4 + i] =
emit_fragment_varying(c,
- decl->Semantic.Name,
- decl->Semantic.Index,
+ semantic_name,
+ semantic_index,
i);
c->num_inputs++;
}
}
-static void
-emit_face_input(struct vc4_compile *c, int attr)
-{
- c->inputs[attr * 4 + 0] = qir_FSUB(c,
- qir_uniform_f(c, 1.0),
- qir_FMUL(c,
- qir_ITOF(c, qir_FRAG_REV_FLAG(c)),
- qir_uniform_f(c, 2.0)));
- c->inputs[attr * 4 + 1] = qir_uniform_f(c, 0.0);
- c->inputs[attr * 4 + 2] = qir_uniform_f(c, 0.0);
- c->inputs[attr * 4 + 3] = qir_uniform_f(c, 1.0);
-}
-
static void
add_output(struct vc4_compile *c,
uint32_t decl_offset,
}
static void
-emit_tgsi_declaration(struct vc4_compile *c,
- struct tgsi_full_declaration *decl)
+declare_uniform_range(struct vc4_compile *c, uint32_t start, uint32_t size)
{
- switch (decl->Declaration.File) {
- case TGSI_FILE_TEMPORARY: {
- uint32_t old_size = c->temps_array_size;
- resize_qreg_array(c, &c->temps, &c->temps_array_size,
- (decl->Range.Last + 1) * 4);
-
- for (int i = old_size; i < c->temps_array_size; i++)
- c->temps[i] = qir_uniform_ui(c, 0);
- break;
+ unsigned array_id = c->num_uniform_ranges++;
+ if (array_id >= c->ubo_ranges_array_size) {
+ c->ubo_ranges_array_size = MAX2(c->ubo_ranges_array_size * 2,
+ array_id + 1);
+ c->ubo_ranges = reralloc(c, c->ubo_ranges,
+ struct vc4_compiler_ubo_range,
+ c->ubo_ranges_array_size);
}
- case TGSI_FILE_INPUT:
- resize_qreg_array(c, &c->inputs, &c->inputs_array_size,
- (decl->Range.Last + 1) * 4);
-
- for (int i = decl->Range.First;
- i <= decl->Range.Last;
- i++) {
- if (c->stage == QSTAGE_FRAG) {
- if (decl->Semantic.Name ==
- TGSI_SEMANTIC_POSITION) {
- emit_fragcoord_input(c, i);
- } else if (decl->Semantic.Name == TGSI_SEMANTIC_FACE) {
- emit_face_input(c, i);
- } else if (decl->Semantic.Name == TGSI_SEMANTIC_GENERIC &&
- (c->fs_key->point_sprite_mask &
- (1 << decl->Semantic.Index))) {
- emit_point_coord_input(c, i);
- } else {
- emit_fragment_input(c, i, decl);
- }
- } else {
- emit_vertex_input(c, i);
- }
- }
- break;
-
- case TGSI_FILE_OUTPUT: {
- for (int i = 0; i < 4; i++) {
- add_output(c,
- decl->Range.First * 4 + i,
- decl->Semantic.Name,
- decl->Semantic.Index,
- i);
- }
-
- switch (decl->Semantic.Name) {
- case TGSI_SEMANTIC_POSITION:
- c->output_position_index = decl->Range.First * 4;
- break;
- case TGSI_SEMANTIC_CLIPVERTEX:
- c->output_clipvertex_index = decl->Range.First * 4;
- break;
- case TGSI_SEMANTIC_COLOR:
- c->output_color_index = decl->Range.First * 4;
- break;
- case TGSI_SEMANTIC_PSIZE:
- c->output_point_size_index = decl->Range.First * 4;
- break;
- }
-
- break;
- }
- }
+ c->ubo_ranges[array_id].dst_offset = 0;
+ c->ubo_ranges[array_id].src_offset = start;
+ c->ubo_ranges[array_id].size = size;
+ c->ubo_ranges[array_id].used = false;
}
static void
-emit_tgsi_instruction(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst)
-{
- struct {
- enum qop op;
- struct qreg (*func)(struct vc4_compile *c,
- struct tgsi_full_instruction *tgsi_inst,
- enum qop op,
- struct qreg *src, int i);
- } op_trans[] = {
- [TGSI_OPCODE_MOV] = { QOP_MOV, tgsi_to_qir_alu },
- [TGSI_OPCODE_ABS] = { 0, tgsi_to_qir_abs },
- [TGSI_OPCODE_MUL] = { QOP_FMUL, tgsi_to_qir_alu },
- [TGSI_OPCODE_ADD] = { QOP_FADD, tgsi_to_qir_alu },
- [TGSI_OPCODE_SUB] = { QOP_FSUB, tgsi_to_qir_alu },
- [TGSI_OPCODE_MIN] = { QOP_FMIN, tgsi_to_qir_alu },
- [TGSI_OPCODE_MAX] = { QOP_FMAX, tgsi_to_qir_alu },
- [TGSI_OPCODE_F2I] = { QOP_FTOI, tgsi_to_qir_alu },
- [TGSI_OPCODE_I2F] = { QOP_ITOF, tgsi_to_qir_alu },
- [TGSI_OPCODE_UADD] = { QOP_ADD, tgsi_to_qir_alu },
- [TGSI_OPCODE_USHR] = { QOP_SHR, tgsi_to_qir_alu },
- [TGSI_OPCODE_ISHR] = { QOP_ASR, tgsi_to_qir_alu },
- [TGSI_OPCODE_SHL] = { QOP_SHL, tgsi_to_qir_alu },
- [TGSI_OPCODE_IMIN] = { QOP_MIN, tgsi_to_qir_alu },
- [TGSI_OPCODE_IMAX] = { QOP_MAX, tgsi_to_qir_alu },
- [TGSI_OPCODE_AND] = { QOP_AND, tgsi_to_qir_alu },
- [TGSI_OPCODE_OR] = { QOP_OR, tgsi_to_qir_alu },
- [TGSI_OPCODE_XOR] = { QOP_XOR, tgsi_to_qir_alu },
- [TGSI_OPCODE_NOT] = { QOP_NOT, tgsi_to_qir_alu },
-
- [TGSI_OPCODE_UMUL] = { 0, tgsi_to_qir_umul },
- [TGSI_OPCODE_IDIV] = { 0, tgsi_to_qir_idiv },
- [TGSI_OPCODE_INEG] = { 0, tgsi_to_qir_ineg },
-
- [TGSI_OPCODE_RSQ] = { QOP_RSQ, tgsi_to_qir_alu },
- [TGSI_OPCODE_SEQ] = { 0, tgsi_to_qir_seq },
- [TGSI_OPCODE_SNE] = { 0, tgsi_to_qir_sne },
- [TGSI_OPCODE_SGE] = { 0, tgsi_to_qir_sge },
- [TGSI_OPCODE_SLT] = { 0, tgsi_to_qir_slt },
- [TGSI_OPCODE_FSEQ] = { 0, tgsi_to_qir_fseq },
- [TGSI_OPCODE_FSNE] = { 0, tgsi_to_qir_fsne },
- [TGSI_OPCODE_FSGE] = { 0, tgsi_to_qir_fsge },
- [TGSI_OPCODE_FSLT] = { 0, tgsi_to_qir_fslt },
- [TGSI_OPCODE_USEQ] = { 0, tgsi_to_qir_useq },
- [TGSI_OPCODE_USNE] = { 0, tgsi_to_qir_usne },
- [TGSI_OPCODE_ISGE] = { 0, tgsi_to_qir_isge },
- [TGSI_OPCODE_ISLT] = { 0, tgsi_to_qir_islt },
-
- [TGSI_OPCODE_CMP] = { 0, tgsi_to_qir_cmp },
- [TGSI_OPCODE_MAD] = { 0, tgsi_to_qir_mad },
- [TGSI_OPCODE_RCP] = { QOP_RCP, tgsi_to_qir_rcp },
- [TGSI_OPCODE_RSQ] = { QOP_RSQ, tgsi_to_qir_scalar },
- [TGSI_OPCODE_EX2] = { QOP_EXP2, tgsi_to_qir_scalar },
- [TGSI_OPCODE_LG2] = { QOP_LOG2, tgsi_to_qir_scalar },
- [TGSI_OPCODE_LRP] = { 0, tgsi_to_qir_lrp },
- [TGSI_OPCODE_TRUNC] = { 0, tgsi_to_qir_trunc },
- [TGSI_OPCODE_CEIL] = { 0, tgsi_to_qir_ceil },
- [TGSI_OPCODE_FRC] = { 0, tgsi_to_qir_frc },
- [TGSI_OPCODE_FLR] = { 0, tgsi_to_qir_flr },
- [TGSI_OPCODE_SIN] = { 0, tgsi_to_qir_sin },
- [TGSI_OPCODE_COS] = { 0, tgsi_to_qir_cos },
- [TGSI_OPCODE_CLAMP] = { 0, tgsi_to_qir_clamp },
- [TGSI_OPCODE_SSG] = { 0, tgsi_to_qir_ssg },
- };
- static int asdf = 0;
- uint32_t tgsi_op = tgsi_inst->Instruction.Opcode;
-
- if (tgsi_op == TGSI_OPCODE_END)
+ntq_emit_alu(struct vc4_compile *c, nir_alu_instr *instr)
+{
+ /* Vectors are special in that they have non-scalarized writemasks,
+ * and just take the first swizzle channel for each argument in order
+ * into each writemask channel.
+ */
+ if (instr->op == nir_op_vec2 ||
+ instr->op == nir_op_vec3 ||
+ instr->op == nir_op_vec4) {
+ struct qreg srcs[4];
+ for (int i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
+ srcs[i] = ntq_get_src(c, instr->src[i].src,
+ instr->src[i].swizzle[0]);
+ struct qreg *dest = ntq_get_dest(c, &instr->dest.dest);
+ for (int i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
+ dest[i] = srcs[i];
return;
-
- struct qreg src_regs[12];
- for (int s = 0; s < 3; s++) {
- for (int i = 0; i < 4; i++) {
- src_regs[4 * s + i] =
- get_src(c, tgsi_inst->Instruction.Opcode,
- &tgsi_inst->Src[s].Register, i);
- }
}
- switch (tgsi_op) {
- case TGSI_OPCODE_TEX:
- case TGSI_OPCODE_TXP:
- case TGSI_OPCODE_TXB:
- case TGSI_OPCODE_TXL:
- tgsi_to_qir_tex(c, tgsi_inst,
- op_trans[tgsi_op].op, src_regs);
- return;
- case TGSI_OPCODE_KILL:
- c->discard = qir_uniform_f(c, 1.0);
- return;
- case TGSI_OPCODE_KILL_IF:
- for (int i = 0; i < 4; i++)
- tgsi_to_qir_kill_if(c, src_regs, i);
- return;
- default:
- break;
+ /* General case: We can just grab the one used channel per src. */
+ struct qreg src[nir_op_infos[instr->op].num_inputs];
+ for (int i = 0; i < nir_op_infos[instr->op].num_inputs; i++) {
+ src[i] = ntq_get_alu_src(c, instr, i);
}
- if (tgsi_op > ARRAY_SIZE(op_trans) || !(op_trans[tgsi_op].func)) {
- fprintf(stderr, "unknown tgsi inst: ");
- tgsi_dump_instruction(tgsi_inst, asdf++);
- fprintf(stderr, "\n");
- abort();
- }
+ /* Pick the channel to store the output in. */
+ assert(!instr->dest.saturate);
+ struct qreg *dest = ntq_get_dest(c, &instr->dest.dest);
+ assert(util_is_power_of_two(instr->dest.write_mask));
+ dest += ffs(instr->dest.write_mask) - 1;
- for (int i = 0; i < 4; i++) {
- if (!(tgsi_inst->Dst[0].Register.WriteMask & (1 << i)))
- continue;
+ switch (instr->op) {
+ case nir_op_fmov:
+ case nir_op_imov:
+ *dest = qir_MOV(c, src[0]);
+ break;
+ case nir_op_fmul:
+ *dest = qir_FMUL(c, src[0], src[1]);
+ break;
+ case nir_op_fadd:
+ *dest = qir_FADD(c, src[0], src[1]);
+ break;
+ case nir_op_fsub:
+ *dest = qir_FSUB(c, src[0], src[1]);
+ break;
+ case nir_op_fmin:
+ *dest = qir_FMIN(c, src[0], src[1]);
+ break;
+ case nir_op_fmax:
+ *dest = qir_FMAX(c, src[0], src[1]);
+ break;
- struct qreg result;
+ case nir_op_f2i:
+ case nir_op_f2u:
+ *dest = qir_FTOI(c, src[0]);
+ break;
+ case nir_op_i2f:
+ case nir_op_u2f:
+ *dest = qir_ITOF(c, src[0]);
+ break;
+ case nir_op_b2f:
+ *dest = qir_AND(c, src[0], qir_uniform_f(c, 1.0));
+ break;
+ case nir_op_b2i:
+ *dest = qir_AND(c, src[0], qir_uniform_ui(c, 1));
+ break;
+ case nir_op_i2b:
+ case nir_op_f2b:
+ qir_SF(c, src[0]);
+ *dest = qir_SEL_X_0_ZC(c, qir_uniform_ui(c, ~0));
+ break;
- result = op_trans[tgsi_op].func(c, tgsi_inst,
- op_trans[tgsi_op].op,
- src_regs, i);
+ case nir_op_iadd:
+ *dest = qir_ADD(c, src[0], src[1]);
+ break;
+ case nir_op_ushr:
+ *dest = qir_SHR(c, src[0], src[1]);
+ break;
+ case nir_op_isub:
+ *dest = qir_SUB(c, src[0], src[1]);
+ break;
+ case nir_op_ishr:
+ *dest = qir_ASR(c, src[0], src[1]);
+ break;
+ case nir_op_ishl:
+ *dest = qir_SHL(c, src[0], src[1]);
+ break;
+ case nir_op_imin:
+ *dest = qir_MIN(c, src[0], src[1]);
+ break;
+ case nir_op_imax:
+ *dest = qir_MAX(c, src[0], src[1]);
+ break;
+ case nir_op_iand:
+ *dest = qir_AND(c, src[0], src[1]);
+ break;
+ case nir_op_ior:
+ *dest = qir_OR(c, src[0], src[1]);
+ break;
+ case nir_op_ixor:
+ *dest = qir_XOR(c, src[0], src[1]);
+ break;
+ case nir_op_inot:
+ *dest = qir_NOT(c, src[0]);
+ break;
- if (tgsi_inst->Instruction.Saturate) {
- float low = (tgsi_inst->Instruction.Saturate ==
- TGSI_SAT_MINUS_PLUS_ONE ? -1.0 : 0.0);
- result = qir_FMAX(c,
- qir_FMIN(c,
- result,
- qir_uniform_f(c, 1.0)),
- qir_uniform_f(c, low));
- }
+ case nir_op_imul:
+ *dest = ntq_umul(c, src[0], src[1]);
+ break;
- update_dst(c, tgsi_inst, i, result);
- }
-}
+ case nir_op_seq:
+ qir_SF(c, qir_FSUB(c, src[0], src[1]));
+ *dest = qir_SEL_X_0_ZS(c, qir_uniform_f(c, 1.0));
+ break;
+ case nir_op_sne:
+ qir_SF(c, qir_FSUB(c, src[0], src[1]));
+ *dest = qir_SEL_X_0_ZC(c, qir_uniform_f(c, 1.0));
+ break;
+ case nir_op_sge:
+ qir_SF(c, qir_FSUB(c, src[0], src[1]));
+ *dest = qir_SEL_X_0_NC(c, qir_uniform_f(c, 1.0));
+ break;
+ case nir_op_slt:
+ qir_SF(c, qir_FSUB(c, src[0], src[1]));
+ *dest = qir_SEL_X_0_NS(c, qir_uniform_f(c, 1.0));
+ break;
+ case nir_op_feq:
+ qir_SF(c, qir_FSUB(c, src[0], src[1]));
+ *dest = qir_SEL_X_0_ZS(c, qir_uniform_ui(c, ~0));
+ break;
+ case nir_op_fne:
+ qir_SF(c, qir_FSUB(c, src[0], src[1]));
+ *dest = qir_SEL_X_0_ZC(c, qir_uniform_ui(c, ~0));
+ break;
+ case nir_op_fge:
+ qir_SF(c, qir_FSUB(c, src[0], src[1]));
+ *dest = qir_SEL_X_0_NC(c, qir_uniform_ui(c, ~0));
+ break;
+ case nir_op_flt:
+ qir_SF(c, qir_FSUB(c, src[0], src[1]));
+ *dest = qir_SEL_X_0_NS(c, qir_uniform_ui(c, ~0));
+ break;
+ case nir_op_ieq:
+ qir_SF(c, qir_SUB(c, src[0], src[1]));
+ *dest = qir_SEL_X_0_ZS(c, qir_uniform_ui(c, ~0));
+ break;
+ case nir_op_ine:
+ qir_SF(c, qir_SUB(c, src[0], src[1]));
+ *dest = qir_SEL_X_0_ZC(c, qir_uniform_ui(c, ~0));
+ break;
+ case nir_op_ige:
+ qir_SF(c, qir_SUB(c, src[0], src[1]));
+ *dest = qir_SEL_X_0_NC(c, qir_uniform_ui(c, ~0));
+ break;
+ case nir_op_ilt:
+ qir_SF(c, qir_SUB(c, src[0], src[1]));
+ *dest = qir_SEL_X_0_NS(c, qir_uniform_ui(c, ~0));
+ break;
-static void
-parse_tgsi_immediate(struct vc4_compile *c, struct tgsi_full_immediate *imm)
-{
- for (int i = 0; i < 4; i++) {
- unsigned n = c->num_consts++;
- resize_qreg_array(c, &c->consts, &c->consts_array_size, n + 1);
- c->consts[n] = qir_uniform_ui(c, imm->u[i].Uint);
+ case nir_op_bcsel:
+ qir_SF(c, src[0]);
+ *dest = qir_SEL_X_Y_NS(c, src[1], src[2]);
+ break;
+ case nir_op_fcsel:
+ qir_SF(c, src[0]);
+ *dest = qir_SEL_X_Y_ZC(c, src[1], src[2]);
+ break;
+
+ case nir_op_frcp:
+ *dest = ntq_rcp(c, src[0]);
+ break;
+ case nir_op_frsq:
+ *dest = ntq_rsq(c, src[0]);
+ break;
+ case nir_op_fexp2:
+ *dest = qir_EXP2(c, src[0]);
+ break;
+ case nir_op_flog2:
+ *dest = qir_LOG2(c, src[0]);
+ break;
+
+ case nir_op_ftrunc:
+ *dest = qir_ITOF(c, qir_FTOI(c, src[0]));
+ break;
+ case nir_op_fceil:
+ *dest = ntq_fceil(c, src[0]);
+ break;
+ case nir_op_ffract:
+ *dest = ntq_ffract(c, src[0]);
+ break;
+ case nir_op_ffloor:
+ *dest = ntq_ffloor(c, src[0]);
+ break;
+
+ case nir_op_fsin:
+ *dest = ntq_fsin(c, src[0]);
+ break;
+ case nir_op_fcos:
+ *dest = ntq_fcos(c, src[0]);
+ break;
+
+ case nir_op_fsign:
+ *dest = ntq_fsign(c, src[0]);
+ break;
+
+ case nir_op_fabs:
+ *dest = qir_FMAXABS(c, src[0], src[0]);
+ break;
+ case nir_op_iabs:
+ *dest = qir_MAX(c, src[0],
+ qir_SUB(c, qir_uniform_ui(c, 0), src[0]));
+ break;
+
+ default:
+ fprintf(stderr, "unknown NIR ALU inst: ");
+ nir_print_instr(&instr->instr, stderr);
+ fprintf(stderr, "\n");
+ abort();
}
}
case PIPE_BLENDFACTOR_DST_COLOR:
return qir_FMUL(c, val, dst[channel]);
case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
- return qir_FMIN(c, src[3], qir_FSUB(c,
- qir_uniform_f(c, 1.0),
- dst[3]));
+ if (channel != 3) {
+ return qir_FMUL(c,
+ val,
+ qir_FMIN(c,
+ src[3],
+ qir_FSUB(c,
+ qir_uniform_f(c, 1.0),
+ dst[3])));
+ } else {
+ return val;
+ }
case PIPE_BLENDFACTOR_CONST_COLOR:
return qir_FMUL(c, val,
- get_temp_for_uniform(c,
- QUNIFORM_BLEND_CONST_COLOR,
- channel));
+ qir_uniform(c, QUNIFORM_BLEND_CONST_COLOR,
+ channel));
case PIPE_BLENDFACTOR_CONST_ALPHA:
return qir_FMUL(c, val,
- get_temp_for_uniform(c,
- QUNIFORM_BLEND_CONST_COLOR,
- 3));
+ qir_uniform(c, QUNIFORM_BLEND_CONST_COLOR, 3));
case PIPE_BLENDFACTOR_ZERO:
return qir_uniform_f(c, 0.0);
case PIPE_BLENDFACTOR_INV_SRC_COLOR:
case PIPE_BLENDFACTOR_INV_CONST_COLOR:
return qir_FMUL(c, val,
qir_FSUB(c, qir_uniform_f(c, 1.0),
- get_temp_for_uniform(c,
- QUNIFORM_BLEND_CONST_COLOR,
- channel)));
+ qir_uniform(c,
+ QUNIFORM_BLEND_CONST_COLOR,
+ channel)));
case PIPE_BLENDFACTOR_INV_CONST_ALPHA:
return qir_FMUL(c, val,
qir_FSUB(c, qir_uniform_f(c, 1.0),
- get_temp_for_uniform(c,
- QUNIFORM_BLEND_CONST_COLOR,
- 3)));
+ qir_uniform(c,
+ QUNIFORM_BLEND_CONST_COLOR,
+ 3)));
default:
case PIPE_BLENDFACTOR_SRC1_COLOR:
return;
}
+ for (int i = 0; i < 4; i++)
+ src_color[i] = qir_SAT(c, src_color[i]);
+
struct qreg src_blend[4], dst_blend[4];
for (int i = 0; i < 3; i++) {
src_blend[i] = vc4_blend_channel(c,
qir_SF(c, dist);
if (c->discard.file == QFILE_NULL)
- c->discard = qir_uniform_f(c, 0.0);
+ c->discard = qir_uniform_ui(c, 0);
- c->discard = qir_SEL_X_Y_NS(c, qir_uniform_f(c, 1.0),
+ c->discard = qir_SEL_X_Y_NS(c, qir_uniform_ui(c, ~0),
c->discard);
}
}
alpha_test_discard(struct vc4_compile *c)
{
struct qreg src_alpha;
- struct qreg alpha_ref = get_temp_for_uniform(c, QUNIFORM_ALPHA_REF, 0);
+ struct qreg alpha_ref = qir_uniform(c, QUNIFORM_ALPHA_REF, 0);
if (!c->fs_key->alpha_test)
return;
src_alpha = qir_uniform_f(c, 1.0);
if (c->discard.file == QFILE_NULL)
- c->discard = qir_uniform_f(c, 0.0);
+ c->discard = qir_uniform_ui(c, 0);
switch (c->fs_key->alpha_test_func) {
case PIPE_FUNC_NEVER:
- c->discard = qir_uniform_f(c, 1.0);
+ c->discard = qir_uniform_ui(c, ~0);
break;
case PIPE_FUNC_ALWAYS:
break;
case PIPE_FUNC_EQUAL:
qir_SF(c, qir_FSUB(c, src_alpha, alpha_ref));
c->discard = qir_SEL_X_Y_ZS(c, c->discard,
- qir_uniform_f(c, 1.0));
+ qir_uniform_ui(c, ~0));
break;
case PIPE_FUNC_NOTEQUAL:
qir_SF(c, qir_FSUB(c, src_alpha, alpha_ref));
c->discard = qir_SEL_X_Y_ZC(c, c->discard,
- qir_uniform_f(c, 1.0));
+ qir_uniform_ui(c, ~0));
break;
case PIPE_FUNC_GREATER:
qir_SF(c, qir_FSUB(c, src_alpha, alpha_ref));
c->discard = qir_SEL_X_Y_NC(c, c->discard,
- qir_uniform_f(c, 1.0));
+ qir_uniform_ui(c, ~0));
break;
case PIPE_FUNC_GEQUAL:
qir_SF(c, qir_FSUB(c, alpha_ref, src_alpha));
c->discard = qir_SEL_X_Y_NS(c, c->discard,
- qir_uniform_f(c, 1.0));
+ qir_uniform_ui(c, ~0));
break;
case PIPE_FUNC_LESS:
qir_SF(c, qir_FSUB(c, src_alpha, alpha_ref));
c->discard = qir_SEL_X_Y_NS(c, c->discard,
- qir_uniform_f(c, 1.0));
+ qir_uniform_ui(c, ~0));
break;
case PIPE_FUNC_LEQUAL:
qir_SF(c, qir_FSUB(c, alpha_ref, src_alpha));
c->discard = qir_SEL_X_Y_NC(c, c->discard,
- qir_uniform_f(c, 1.0));
+ qir_uniform_ui(c, ~0));
break;
}
}
-static void
-emit_frag_end(struct vc4_compile *c)
+static struct qreg
+vc4_logicop(struct vc4_compile *c, struct qreg src, struct qreg dst)
{
- clip_distance_discard(c);
- alpha_test_discard(c);
+ switch (c->fs_key->logicop_func) {
+ case PIPE_LOGICOP_CLEAR:
+ return qir_uniform_f(c, 0.0);
+ case PIPE_LOGICOP_NOR:
+ return qir_NOT(c, qir_OR(c, src, dst));
+ case PIPE_LOGICOP_AND_INVERTED:
+ return qir_AND(c, qir_NOT(c, src), dst);
+ case PIPE_LOGICOP_COPY_INVERTED:
+ return qir_NOT(c, src);
+ case PIPE_LOGICOP_AND_REVERSE:
+ return qir_AND(c, src, qir_NOT(c, dst));
+ case PIPE_LOGICOP_INVERT:
+ return qir_NOT(c, dst);
+ case PIPE_LOGICOP_XOR:
+ return qir_XOR(c, src, dst);
+ case PIPE_LOGICOP_NAND:
+ return qir_NOT(c, qir_AND(c, src, dst));
+ case PIPE_LOGICOP_AND:
+ return qir_AND(c, src, dst);
+ case PIPE_LOGICOP_EQUIV:
+ return qir_NOT(c, qir_XOR(c, src, dst));
+ case PIPE_LOGICOP_NOOP:
+ return dst;
+ case PIPE_LOGICOP_OR_INVERTED:
+ return qir_OR(c, qir_NOT(c, src), dst);
+ case PIPE_LOGICOP_OR_REVERSE:
+ return qir_OR(c, src, qir_NOT(c, dst));
+ case PIPE_LOGICOP_OR:
+ return qir_OR(c, src, dst);
+ case PIPE_LOGICOP_SET:
+ return qir_uniform_ui(c, ~0);
+ case PIPE_LOGICOP_COPY:
+ default:
+ return src;
+ }
+}
+/**
+ * Applies the GL blending pipeline and returns the packed (8888) output
+ * color.
+ */
+static struct qreg
+blend_pipeline(struct vc4_compile *c)
+{
enum pipe_format color_format = c->fs_key->color_format;
const uint8_t *format_swiz = vc4_get_format_swizzle(color_format);
struct qreg tlb_read_color[4] = { c->undef, c->undef, c->undef, c->undef };
struct qreg dst_color[4] = { c->undef, c->undef, c->undef, c->undef };
struct qreg linear_dst_color[4] = { c->undef, c->undef, c->undef, c->undef };
+ struct qreg packed_dst_color = c->undef;
+
if (c->fs_key->blend.blend_enable ||
- c->fs_key->blend.colormask != 0xf) {
- struct qreg r4 = qir_TLB_COLOR_READ(c);
+ c->fs_key->blend.colormask != 0xf ||
+ c->fs_key->logicop_func != PIPE_LOGICOP_COPY) {
+ packed_dst_color = qir_TLB_COLOR_READ(c);
for (int i = 0; i < 4; i++)
- tlb_read_color[i] = qir_R4_UNPACK(c, r4, i);
+ tlb_read_color[i] = qir_UNPACK_8_F(c,
+ packed_dst_color, i);
for (int i = 0; i < 4; i++) {
dst_color[i] = get_swizzled_channel(c,
tlb_read_color,
}
}
+ struct qreg undef_array[4] = { c->undef, c->undef, c->undef, c->undef };
+ const struct qreg *output_colors = (c->output_color_index != -1 ?
+ c->outputs + c->output_color_index :
+ undef_array);
+ struct qreg blend_src_color[4];
+ for (int i = 0; i < 4; i++)
+ blend_src_color[i] = output_colors[i];
+
struct qreg blend_color[4];
- struct qreg undef_array[4] = {
- c->undef, c->undef, c->undef, c->undef
- };
- vc4_blend(c, blend_color, linear_dst_color,
- (c->output_color_index != -1 ?
- c->outputs + c->output_color_index :
- undef_array));
+ vc4_blend(c, blend_color, linear_dst_color, blend_src_color);
if (util_format_is_srgb(color_format)) {
for (int i = 0; i < 3; i++)
blend_color[i] = qir_srgb_encode(c, blend_color[i]);
}
- /* If the bit isn't set in the color mask, then just return the
- * original dst color, instead.
- */
- for (int i = 0; i < 4; i++) {
- if (!(c->fs_key->blend.colormask & (1 << i))) {
- blend_color[i] = dst_color[i];
- }
- }
-
/* Debug: Sometimes you're getting a black output and just want to see
* if the FS is getting executed at all. Spam magenta into the color
* output.
format_swiz[i]);
}
+ struct qreg packed_color = c->undef;
+ for (int i = 0; i < 4; i++) {
+ if (swizzled_outputs[i].file == QFILE_NULL)
+ continue;
+ if (packed_color.file == QFILE_NULL) {
+ packed_color = qir_PACK_8888_F(c, swizzled_outputs[i]);
+ } else {
+ packed_color = qir_PACK_8_F(c,
+ packed_color,
+ swizzled_outputs[i],
+ i);
+ }
+ }
+
+ if (packed_color.file == QFILE_NULL)
+ packed_color = qir_uniform_ui(c, 0);
+
+ if (c->fs_key->logicop_func != PIPE_LOGICOP_COPY) {
+ packed_color = vc4_logicop(c, packed_color, packed_dst_color);
+ }
+
+ /* If the bit isn't set in the color mask, then just return the
+ * original dst color, instead.
+ */
+ uint32_t colormask = 0xffffffff;
+ for (int i = 0; i < 4; i++) {
+ if (format_swiz[i] < 4 &&
+ !(c->fs_key->blend.colormask & (1 << format_swiz[i]))) {
+ colormask &= ~(0xff << (i * 8));
+ }
+ }
+ if (colormask != 0xffffffff) {
+ packed_color = qir_OR(c,
+ qir_AND(c, packed_color,
+ qir_uniform_ui(c, colormask)),
+ qir_AND(c, packed_dst_color,
+ qir_uniform_ui(c, ~colormask)));
+ }
+
+ return packed_color;
+}
+
+static void
+emit_frag_end(struct vc4_compile *c)
+{
+ clip_distance_discard(c);
+ alpha_test_discard(c);
+ struct qreg color = blend_pipeline(c);
+
if (c->discard.file != QFILE_NULL)
qir_TLB_DISCARD_SETUP(c, c->discard);
if (c->fs_key->stencil_enabled) {
- qir_TLB_STENCIL_SETUP(c, add_uniform(c, QUNIFORM_STENCIL, 0));
+ qir_TLB_STENCIL_SETUP(c, qir_uniform(c, QUNIFORM_STENCIL, 0));
if (c->fs_key->stencil_twoside) {
- qir_TLB_STENCIL_SETUP(c, add_uniform(c, QUNIFORM_STENCIL, 1));
+ qir_TLB_STENCIL_SETUP(c, qir_uniform(c, QUNIFORM_STENCIL, 1));
}
if (c->fs_key->stencil_full_writemasks) {
- qir_TLB_STENCIL_SETUP(c, add_uniform(c, QUNIFORM_STENCIL, 2));
+ qir_TLB_STENCIL_SETUP(c, qir_uniform(c, QUNIFORM_STENCIL, 2));
}
}
qir_TLB_Z_WRITE(c, z);
}
- bool color_written = false;
- for (int i = 0; i < 4; i++) {
- if (swizzled_outputs[i].file != QFILE_NULL)
- color_written = true;
- }
-
- struct qreg packed_color;
- if (color_written) {
- /* Fill in any undefined colors. The simulator will assertion
- * fail if we read something that wasn't written, and I don't
- * know what hardware does.
- */
- for (int i = 0; i < 4; i++) {
- if (swizzled_outputs[i].file == QFILE_NULL)
- swizzled_outputs[i] = qir_uniform_f(c, 0.0);
- }
- packed_color = qir_get_temp(c);
- qir_emit(c, qir_inst4(QOP_PACK_COLORS, packed_color,
- swizzled_outputs[0],
- swizzled_outputs[1],
- swizzled_outputs[2],
- swizzled_outputs[3]));
- } else {
- packed_color = qir_uniform_ui(c, 0);
- }
-
- qir_emit(c, qir_inst(QOP_TLB_COLOR_WRITE, c->undef,
- packed_color, c->undef));
+ qir_TLB_COLOR_WRITE(c, color);
}
static void
for (int i = 0; i < 2; i++) {
struct qreg scale =
- add_uniform(c, QUNIFORM_VIEWPORT_X_SCALE + i, 0);
+ qir_uniform(c, QUNIFORM_VIEWPORT_X_SCALE + i, 0);
xyi[i] = qir_FTOI(c, qir_FMUL(c,
qir_FMUL(c,
- c->outputs[i],
+ c->outputs[c->output_position_index + i],
scale),
rcp_w));
}
static void
emit_zs_write(struct vc4_compile *c, struct qreg rcp_w)
{
- struct qreg zscale = add_uniform(c, QUNIFORM_VIEWPORT_Z_SCALE, 0);
- struct qreg zoffset = add_uniform(c, QUNIFORM_VIEWPORT_Z_OFFSET, 0);
+ struct qreg zscale = qir_uniform(c, QUNIFORM_VIEWPORT_Z_SCALE, 0);
+ struct qreg zoffset = qir_uniform(c, QUNIFORM_VIEWPORT_Z_OFFSET, 0);
- qir_VPM_WRITE(c, qir_FMUL(c, qir_FADD(c, qir_FMUL(c,
- c->outputs[2],
+ qir_VPM_WRITE(c, qir_FADD(c, qir_FMUL(c, qir_FMUL(c,
+ c->outputs[c->output_position_index + 2],
zscale),
- zoffset),
- rcp_w));
+ rcp_w),
+ zoffset));
}
static void
{
struct qreg point_size;
- if (c->output_point_size_index)
+ if (c->output_point_size_index != -1)
point_size = c->outputs[c->output_point_size_index + 3];
else
point_size = qir_uniform_f(c, 1.0);
if (c->num_inputs)
return;
- for (int i = 0; i < 4; i++) {
- qir_emit(c, qir_inst(QOP_VPM_READ,
- qir_get_temp(c),
- c->undef,
- c->undef));
- c->num_inputs++;
- }
+ c->vattr_sizes[0] = 4;
+ struct qreg vpm = { QFILE_VPM, 0 };
+ (void)qir_MOV(c, vpm);
+ c->num_inputs++;
}
static void
emit_ucp_clipdistance(struct vc4_compile *c)
{
- struct qreg *clipvertex;
-
+ unsigned cv;
if (c->output_clipvertex_index != -1)
- clipvertex = &c->outputs[c->output_clipvertex_index];
+ cv = c->output_clipvertex_index;
else if (c->output_position_index != -1)
- clipvertex = &c->outputs[c->output_position_index];
+ cv = c->output_position_index;
else
return;
plane,
TGSI_SWIZZLE_X);
+
struct qreg dist = qir_uniform_f(c, 0.0);
for (int i = 0; i < 4; i++) {
+ struct qreg pos_chan = c->outputs[cv + i];
struct qreg ucp =
- add_uniform(c, QUNIFORM_USER_CLIP_PLANE,
+ qir_uniform(c, QUNIFORM_USER_CLIP_PLANE,
plane * 4 + i);
- dist = qir_FADD(c, dist, qir_FMUL(c, clipvertex[i], ucp));
+ dist = qir_FADD(c, dist, qir_FMUL(c, pos_chan, ucp));
}
c->outputs[output_index] = dist;
struct vc4_varying_semantic *fs_inputs,
uint32_t num_fs_inputs)
{
- struct qreg rcp_w = qir_RCP(c, c->outputs[3]);
+ struct qreg rcp_w = qir_RCP(c, c->outputs[c->output_position_index + 3]);
emit_stub_vpm_read(c);
emit_ucp_clipdistance(c);
static void
emit_coord_end(struct vc4_compile *c)
{
- struct qreg rcp_w = qir_RCP(c, c->outputs[3]);
+ struct qreg rcp_w = qir_RCP(c, c->outputs[c->output_position_index + 3]);
emit_stub_vpm_read(c);
for (int i = 0; i < 4; i++)
- qir_VPM_WRITE(c, c->outputs[i]);
+ qir_VPM_WRITE(c, c->outputs[c->output_position_index + i]);
emit_scaled_viewport_write(c, rcp_w);
emit_zs_write(c, rcp_w);
emit_point_size_write(c);
}
+static void
+vc4_optimize_nir(struct nir_shader *s)
+{
+ bool progress;
+
+ do {
+ progress = false;
+
+ nir_lower_vars_to_ssa(s);
+ nir_lower_alu_to_scalar(s);
+
+ progress = nir_copy_prop(s) || progress;
+ progress = nir_opt_dce(s) || progress;
+ progress = nir_opt_cse(s) || progress;
+ progress = nir_opt_peephole_select(s) || progress;
+ progress = nir_opt_algebraic(s) || progress;
+ progress = nir_opt_constant_folding(s) || progress;
+ } while (progress);
+}
+
+static int
+driver_location_compare(const void *in_a, const void *in_b)
+{
+ const nir_variable *const *a = in_a;
+ const nir_variable *const *b = in_b;
+
+ return (*a)->data.driver_location - (*b)->data.driver_location;
+}
+
+static void
+ntq_setup_inputs(struct vc4_compile *c)
+{
+ unsigned num_entries = 0;
+ foreach_list_typed(nir_variable, var, node, &c->s->inputs)
+ num_entries++;
+
+ nir_variable *vars[num_entries];
+
+ unsigned i = 0;
+ foreach_list_typed(nir_variable, var, node, &c->s->inputs)
+ vars[i++] = var;
+
+ /* Sort the variables so that we emit the input setup in
+ * driver_location order. This is required for VPM reads, whose data
+ * is fetched into the VPM in driver_location (TGSI register index)
+ * order.
+ */
+ qsort(&vars, num_entries, sizeof(*vars), driver_location_compare);
+
+ for (unsigned i = 0; i < num_entries; i++) {
+ nir_variable *var = vars[i];
+ unsigned array_len = MAX2(glsl_get_length(var->type), 1);
+ /* XXX: map loc slots to semantics */
+ unsigned semantic_name = var->data.location;
+ unsigned semantic_index = var->data.index;
+ unsigned loc = var->data.driver_location;
+
+ assert(array_len == 1);
+ (void)array_len;
+ resize_qreg_array(c, &c->inputs, &c->inputs_array_size,
+ (loc + 1) * 4);
+
+ if (c->stage == QSTAGE_FRAG) {
+ if (semantic_name == TGSI_SEMANTIC_POSITION) {
+ emit_fragcoord_input(c, loc);
+ } else if (semantic_name == TGSI_SEMANTIC_FACE) {
+ c->inputs[loc * 4 + 0] = qir_FRAG_REV_FLAG(c);
+ } else if (semantic_name == TGSI_SEMANTIC_GENERIC &&
+ (c->fs_key->point_sprite_mask &
+ (1 << semantic_index))) {
+ c->inputs[loc * 4 + 0] = c->point_x;
+ c->inputs[loc * 4 + 1] = c->point_y;
+ } else {
+ emit_fragment_input(c, loc,
+ semantic_name,
+ semantic_index);
+ }
+ } else {
+ emit_vertex_input(c, loc);
+ }
+ }
+}
+
+static void
+ntq_setup_outputs(struct vc4_compile *c)
+{
+ foreach_list_typed(nir_variable, var, node, &c->s->outputs) {
+ unsigned array_len = MAX2(glsl_get_length(var->type), 1);
+ /* XXX: map loc slots to semantics */
+ unsigned semantic_name = var->data.location;
+ unsigned semantic_index = var->data.index;
+ unsigned loc = var->data.driver_location * 4;
+
+ assert(array_len == 1);
+ (void)array_len;
+
+ /* NIR hack to pass through
+ * TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS */
+ if (semantic_name == TGSI_SEMANTIC_COLOR &&
+ semantic_index == -1)
+ semantic_index = 0;
+
+ for (int i = 0; i < 4; i++) {
+ add_output(c,
+ loc + i,
+ semantic_name,
+ semantic_index,
+ i);
+ }
+
+ switch (semantic_name) {
+ case TGSI_SEMANTIC_POSITION:
+ c->output_position_index = loc;
+ break;
+ case TGSI_SEMANTIC_CLIPVERTEX:
+ c->output_clipvertex_index = loc;
+ break;
+ case TGSI_SEMANTIC_COLOR:
+ c->output_color_index = loc;
+ break;
+ case TGSI_SEMANTIC_PSIZE:
+ c->output_point_size_index = loc;
+ break;
+ }
+
+ }
+}
+
+static void
+ntq_setup_uniforms(struct vc4_compile *c)
+{
+ foreach_list_typed(nir_variable, var, node, &c->s->uniforms) {
+ unsigned array_len = MAX2(glsl_get_length(var->type), 1);
+ unsigned array_elem_size = 4 * sizeof(float);
+
+ declare_uniform_range(c, var->data.driver_location * array_elem_size,
+ array_len * array_elem_size);
+
+ }
+}
+
+/**
+ * Sets up the mapping from nir_register to struct qreg *.
+ *
+ * Each nir_register gets a struct qreg per 32-bit component being stored.
+ */
+static void
+ntq_setup_registers(struct vc4_compile *c, struct exec_list *list)
+{
+ foreach_list_typed(nir_register, nir_reg, node, list) {
+ unsigned array_len = MAX2(nir_reg->num_array_elems, 1);
+ struct qreg *qregs = ralloc_array(c->def_ht, struct qreg,
+ array_len *
+ nir_reg->num_components);
+
+ _mesa_hash_table_insert(c->def_ht, nir_reg, qregs);
+
+ for (int i = 0; i < array_len * nir_reg->num_components; i++)
+ qregs[i] = qir_uniform_ui(c, 0);
+ }
+}
+
+static void
+ntq_emit_load_const(struct vc4_compile *c, nir_load_const_instr *instr)
+{
+ struct qreg *qregs = ntq_init_ssa_def(c, &instr->def);
+ for (int i = 0; i < instr->def.num_components; i++)
+ qregs[i] = qir_uniform_ui(c, instr->value.u[i]);
+
+ _mesa_hash_table_insert(c->def_ht, &instr->def, qregs);
+}
+
+static void
+ntq_emit_ssa_undef(struct vc4_compile *c, nir_ssa_undef_instr *instr)
+{
+ struct qreg *qregs = ntq_init_ssa_def(c, &instr->def);
+
+ /* QIR needs there to be *some* value, so pick 0 (same as for
+ * ntq_setup_registers().
+ */
+ for (int i = 0; i < instr->def.num_components; i++)
+ qregs[i] = qir_uniform_ui(c, 0);
+}
+
+static void
+ntq_emit_intrinsic(struct vc4_compile *c, nir_intrinsic_instr *instr)
+{
+ const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic];
+ struct qreg *dest = NULL;
+
+ if (info->has_dest) {
+ dest = ntq_get_dest(c, &instr->dest);
+ }
+
+ switch (instr->intrinsic) {
+ case nir_intrinsic_load_uniform:
+ assert(instr->num_components == 1);
+ *dest = qir_uniform(c, QUNIFORM_UNIFORM, instr->const_index[0]);
+ break;
+
+ case nir_intrinsic_load_uniform_indirect:
+ *dest = indirect_uniform_load(c, instr);
+
+ break;
+
+ case nir_intrinsic_load_input:
+ assert(instr->num_components == 1);
+ *dest = c->inputs[instr->const_index[0]];
+
+ break;
+
+ case nir_intrinsic_store_output:
+ assert(instr->num_components == 1);
+ c->outputs[instr->const_index[0]] =
+ qir_MOV(c, ntq_get_src(c, instr->src[0], 0));
+ c->num_outputs = MAX2(c->num_outputs, instr->const_index[0] + 1);
+ break;
+
+ case nir_intrinsic_discard:
+ c->discard = qir_uniform_ui(c, ~0);
+ break;
+
+ case nir_intrinsic_discard_if:
+ if (c->discard.file == QFILE_NULL)
+ c->discard = qir_uniform_ui(c, 0);
+ c->discard = qir_OR(c, c->discard,
+ ntq_get_src(c, instr->src[0], 0));
+ break;
+
+ default:
+ fprintf(stderr, "Unknown intrinsic: ");
+ nir_print_instr(&instr->instr, stderr);
+ fprintf(stderr, "\n");
+ break;
+ }
+}
+
+static void
+ntq_emit_if(struct vc4_compile *c, nir_if *if_stmt)
+{
+ fprintf(stderr, "general IF statements not handled.\n");
+}
+
+static void
+ntq_emit_instr(struct vc4_compile *c, nir_instr *instr)
+{
+ switch (instr->type) {
+ case nir_instr_type_alu:
+ ntq_emit_alu(c, nir_instr_as_alu(instr));
+ break;
+
+ case nir_instr_type_intrinsic:
+ ntq_emit_intrinsic(c, nir_instr_as_intrinsic(instr));
+ break;
+
+ case nir_instr_type_load_const:
+ ntq_emit_load_const(c, nir_instr_as_load_const(instr));
+ break;
+
+ case nir_instr_type_ssa_undef:
+ ntq_emit_ssa_undef(c, nir_instr_as_ssa_undef(instr));
+ break;
+
+ case nir_instr_type_tex:
+ ntq_emit_tex(c, nir_instr_as_tex(instr));
+ break;
+
+ default:
+ fprintf(stderr, "Unknown NIR instr type: ");
+ nir_print_instr(instr, stderr);
+ fprintf(stderr, "\n");
+ abort();
+ }
+}
+
+static void
+ntq_emit_block(struct vc4_compile *c, nir_block *block)
+{
+ nir_foreach_instr(block, instr) {
+ ntq_emit_instr(c, instr);
+ }
+}
+
+static void
+ntq_emit_cf_list(struct vc4_compile *c, struct exec_list *list)
+{
+ foreach_list_typed(nir_cf_node, node, node, list) {
+ switch (node->type) {
+ /* case nir_cf_node_loop: */
+ case nir_cf_node_block:
+ ntq_emit_block(c, nir_cf_node_as_block(node));
+ break;
+
+ case nir_cf_node_if:
+ ntq_emit_if(c, nir_cf_node_as_if(node));
+ break;
+
+ default:
+ assert(0);
+ }
+ }
+}
+
+static void
+ntq_emit_impl(struct vc4_compile *c, nir_function_impl *impl)
+{
+ ntq_setup_registers(c, &impl->registers);
+ ntq_emit_cf_list(c, &impl->body);
+}
+
+static void
+nir_to_qir(struct vc4_compile *c)
+{
+ ntq_setup_inputs(c);
+ ntq_setup_outputs(c);
+ ntq_setup_uniforms(c);
+ ntq_setup_registers(c, &c->s->registers);
+
+ /* Find the main function and emit the body. */
+ nir_foreach_overload(c->s, overload) {
+ assert(strcmp(overload->function->name, "main") == 0);
+ assert(overload->impl);
+ ntq_emit_impl(c, overload->impl);
+ }
+}
+
+static const nir_shader_compiler_options nir_options = {
+ .lower_ffma = true,
+ .lower_flrp = true,
+ .lower_fpow = true,
+ .lower_fsat = true,
+ .lower_fsqrt = true,
+ .lower_negate = true,
+};
+
+static bool
+count_nir_instrs_in_block(nir_block *block, void *state)
+{
+ int *count = (int *) state;
+ nir_foreach_instr(block, instr) {
+ *count = *count + 1;
+ }
+ return true;
+}
+
+static int
+count_nir_instrs(nir_shader *nir)
+{
+ int count = 0;
+ nir_foreach_overload(nir, overload) {
+ if (!overload->impl)
+ continue;
+ nir_foreach_block(overload->impl, count_nir_instrs_in_block, &count);
+ }
+ return count;
+}
+
static struct vc4_compile *
-vc4_shader_tgsi_to_qir(struct vc4_context *vc4, enum qstage stage,
+vc4_shader_ntq(struct vc4_context *vc4, enum qstage stage,
struct vc4_key *key)
{
struct vc4_compile *c = qir_compile_init();
- int ret;
c->stage = stage;
c->shader_state = &key->shader_state->base;
+ c->program_id = key->shader_state->program_id;
+ c->variant_id = key->shader_state->compiled_variant_count++;
c->key = key;
switch (stage) {
tokens = key->shader_state->twoside_tokens;
}
- ret = tgsi_parse_init(&c->parser, tokens);
- assert(ret == TGSI_PARSE_OK);
-
if (vc4_debug & VC4_DEBUG_TGSI) {
- fprintf(stderr, "TGSI:\n");
+ fprintf(stderr, "%s prog %d/%d TGSI:\n",
+ qir_get_stage_name(c->stage),
+ c->program_id, c->variant_id);
tgsi_dump(tokens, 0);
}
- while (!tgsi_parse_end_of_tokens(&c->parser)) {
- tgsi_parse_token(&c->parser);
+ c->s = tgsi_to_nir(tokens, &nir_options);
+ nir_opt_global_to_local(c->s);
+ nir_convert_to_ssa(c->s);
+ vc4_nir_lower_io(c);
+ nir_lower_idiv(c->s);
+ nir_lower_load_const_to_scalar(c->s);
- switch (c->parser.FullToken.Token.Type) {
- case TGSI_TOKEN_TYPE_DECLARATION:
- emit_tgsi_declaration(c,
- &c->parser.FullToken.FullDeclaration);
- break;
+ vc4_optimize_nir(c->s);
- case TGSI_TOKEN_TYPE_INSTRUCTION:
- emit_tgsi_instruction(c,
- &c->parser.FullToken.FullInstruction);
- break;
+ nir_remove_dead_variables(c->s);
- case TGSI_TOKEN_TYPE_IMMEDIATE:
- parse_tgsi_immediate(c,
- &c->parser.FullToken.FullImmediate);
- break;
- }
+ nir_convert_from_ssa(c->s, true);
+
+ if (vc4_debug & VC4_DEBUG_SHADERDB) {
+ fprintf(stderr, "SHADER-DB: %s prog %d/%d: %d NIR instructions\n",
+ qir_get_stage_name(c->stage),
+ c->program_id, c->variant_id,
+ count_nir_instrs(c->s));
+ }
+
+ if (vc4_debug & VC4_DEBUG_NIR) {
+ fprintf(stderr, "%s prog %d/%d NIR:\n",
+ qir_get_stage_name(c->stage),
+ c->program_id, c->variant_id);
+ nir_print_shader(c->s, stderr);
}
+ nir_to_qir(c);
+
switch (stage) {
case QSTAGE_FRAG:
emit_frag_end(c);
break;
}
- tgsi_parse_free(&c->parser);
+ if (vc4_debug & VC4_DEBUG_QIR) {
+ fprintf(stderr, "%s prog %d/%d pre-opt QIR:\n",
+ qir_get_stage_name(c->stage),
+ c->program_id, c->variant_id);
+ qir_dump(c);
+ }
qir_optimize(c);
+ qir_lower_uniforms(c);
if (vc4_debug & VC4_DEBUG_QIR) {
- fprintf(stderr, "QIR:\n");
+ fprintf(stderr, "%s prog %d/%d QIR:\n",
+ qir_get_stage_name(c->stage),
+ c->program_id, c->variant_id);
qir_dump(c);
}
qir_reorder_uniforms(c);
vc4_generate_code(vc4, c);
if (vc4_debug & VC4_DEBUG_SHADERDB) {
- fprintf(stderr, "SHADER-DB: %s: %d instructions\n",
- qir_get_stage_name(c->stage), c->qpu_inst_count);
- fprintf(stderr, "SHADER-DB: %s: %d uniforms\n",
- qir_get_stage_name(c->stage), c->num_uniforms);
+ fprintf(stderr, "SHADER-DB: %s prog %d/%d: %d instructions\n",
+ qir_get_stage_name(c->stage),
+ c->program_id, c->variant_id,
+ c->qpu_inst_count);
+ fprintf(stderr, "SHADER-DB: %s prog %d/%d: %d uniforms\n",
+ qir_get_stage_name(c->stage),
+ c->program_id, c->variant_id,
+ c->num_uniforms);
}
+ ralloc_free(c->s);
+
return c;
}
vc4_shader_state_create(struct pipe_context *pctx,
const struct pipe_shader_state *cso)
{
+ struct vc4_context *vc4 = vc4_context(pctx);
struct vc4_uncompiled_shader *so = CALLOC_STRUCT(vc4_uncompiled_shader);
if (!so)
return NULL;
- const struct tgsi_lowering_config lowering_config = {
- .lower_DST = true,
- .lower_XPD = true,
- .lower_SCS = true,
- .lower_POW = true,
- .lower_LIT = true,
- .lower_EXP = true,
- .lower_LOG = true,
- .lower_DP4 = true,
- .lower_DP3 = true,
- .lower_DPH = true,
- .lower_DP2 = true,
- .lower_DP2A = true,
- };
-
- struct tgsi_shader_info info;
- so->base.tokens = tgsi_transform_lowering(&lowering_config, cso->tokens, &info);
- if (!so->base.tokens)
- so->base.tokens = tgsi_dup_tokens(cso->tokens);
+ so->base.tokens = tgsi_dup_tokens(cso->tokens);
+ so->program_id = vc4->next_uncompiled_program_id++;
return so;
}
memcpy(uinfo->contents, c->uniform_contents,
count * sizeof(*uinfo->contents));
uinfo->num_texture_samples = c->num_texture_samples;
+
+ vc4_set_shader_uniform_dirty_flags(shader);
}
static struct vc4_compiled_shader *
vc4_get_compiled_shader(struct vc4_context *vc4, enum qstage stage,
struct vc4_key *key)
{
- struct util_hash_table *ht;
+ struct hash_table *ht;
uint32_t key_size;
if (stage == QSTAGE_FRAG) {
ht = vc4->fs_cache;
}
struct vc4_compiled_shader *shader;
- shader = util_hash_table_get(ht, key);
- if (shader)
- return shader;
+ struct hash_entry *entry = _mesa_hash_table_search(ht, key);
+ if (entry)
+ return entry->data;
- struct vc4_compile *c = vc4_shader_tgsi_to_qir(vc4, stage, key);
+ struct vc4_compile *c = vc4_shader_ntq(vc4, stage, key);
shader = rzalloc(NULL, struct vc4_compiled_shader);
shader->program_id = vc4->next_compiled_program_id++;
if (stage == QSTAGE_FRAG) {
+ bool input_live[c->num_input_semantics];
+
+ memset(input_live, 0, sizeof(input_live));
+ list_for_each_entry(struct qinst, inst, &c->instructions, link) {
+ for (int i = 0; i < qir_get_op_nsrc(inst->op); i++) {
+ if (inst->src[i].file == QFILE_VARY)
+ input_live[inst->src[i].index] = true;
+ }
+ }
+
shader->input_semantics = ralloc_array(shader,
struct vc4_varying_semantic,
c->num_input_semantics);
for (int i = 0; i < c->num_input_semantics; i++) {
struct vc4_varying_semantic *sem = &c->input_semantics[i];
+ if (!input_live[i])
+ continue;
+
/* Skip non-VS-output inputs. */
if (sem->semantic == (uint8_t)~0)
continue;
- if (sem->semantic == TGSI_SEMANTIC_COLOR)
+ if (sem->semantic == TGSI_SEMANTIC_COLOR ||
+ sem->semantic == TGSI_SEMANTIC_BCOLOR) {
shader->color_inputs |= (1 << shader->num_inputs);
+ }
+
shader->input_semantics[shader->num_inputs] = *sem;
shader->num_inputs++;
}
} else {
shader->num_inputs = c->num_inputs;
+
+ shader->vattr_offsets[0] = 0;
+ for (int i = 0; i < 8; i++) {
+ shader->vattr_offsets[i + 1] =
+ shader->vattr_offsets[i] + c->vattr_sizes[i];
+
+ if (c->vattr_sizes[i])
+ shader->vattrs_live |= (1 << i);
+ }
}
copy_uniform_state_to_shader(shader, c);
- shader->bo = vc4_bo_alloc_mem(vc4->screen, c->qpu_insts,
- c->qpu_inst_count * sizeof(uint64_t),
- "code");
+ shader->bo = vc4_bo_alloc_shader(vc4->screen, c->qpu_insts,
+ c->qpu_inst_count * sizeof(uint64_t));
+
+ /* Copy the compiler UBO range state to the compiled shader, dropping
+ * out arrays that were never referenced by an indirect load.
+ *
+ * (Note that QIR dead code elimination of an array access still
+ * leaves that array alive, though)
+ */
+ if (c->num_ubo_ranges) {
+ shader->num_ubo_ranges = c->num_ubo_ranges;
+ shader->ubo_ranges = ralloc_array(shader, struct vc4_ubo_range,
+ c->num_ubo_ranges);
+ uint32_t j = 0;
+ for (int i = 0; i < c->num_uniform_ranges; i++) {
+ struct vc4_compiler_ubo_range *range =
+ &c->ubo_ranges[i];
+ if (!range->used)
+ continue;
+
+ shader->ubo_ranges[j].dst_offset = range->dst_offset;
+ shader->ubo_ranges[j].src_offset = range->src_offset;
+ shader->ubo_ranges[j].size = range->size;
+ shader->ubo_size += c->ubo_ranges[i].size;
+ j++;
+ }
+ }
+ if (shader->ubo_size) {
+ if (vc4_debug & VC4_DEBUG_SHADERDB) {
+ fprintf(stderr, "SHADER-DB: %s prog %d/%d: %d UBO uniforms\n",
+ qir_get_stage_name(c->stage),
+ c->program_id, c->variant_id,
+ shader->ubo_size / 4);
+ }
+ }
qir_compile_destroy(c);
struct vc4_key *dup_key;
- dup_key = malloc(key_size);
+ dup_key = ralloc_size(shader, key_size);
memcpy(dup_key, key, key_size);
- util_hash_table_set(ht, dup_key, shader);
+ _mesa_hash_table_insert(ht, dup_key, shader);
return shader;
}
VC4_DIRTY_RASTERIZER |
VC4_DIRTY_FRAGTEX |
VC4_DIRTY_TEXSTATE |
- VC4_DIRTY_PROG))) {
+ VC4_DIRTY_UNCOMPILED_FS))) {
return;
}
key->is_lines = (prim_mode >= PIPE_PRIM_LINES &&
prim_mode <= PIPE_PRIM_LINE_STRIP);
key->blend = vc4->blend->rt[0];
-
+ if (vc4->blend->logicop_enable) {
+ key->logicop_func = vc4->blend->logicop_func;
+ } else {
+ key->logicop_func = PIPE_LOGICOP_COPY;
+ }
if (vc4->framebuffer.cbufs[0])
key->color_format = vc4->framebuffer.cbufs[0]->format;
if (vc4->prog.fs == old_fs)
return;
+ vc4->dirty |= VC4_DIRTY_COMPILED_FS;
if (vc4->rasterizer->base.flatshade &&
old_fs && vc4->prog.fs->color_inputs != old_fs->color_inputs) {
vc4->dirty |= VC4_DIRTY_FLAT_SHADE_FLAGS;
VC4_DIRTY_VERTTEX |
VC4_DIRTY_TEXSTATE |
VC4_DIRTY_VTXSTATE |
- VC4_DIRTY_PROG))) {
+ VC4_DIRTY_UNCOMPILED_VS |
+ VC4_DIRTY_COMPILED_FS))) {
return;
}
(prim_mode == PIPE_PRIM_POINTS &&
vc4->rasterizer->base.point_size_per_vertex);
- vc4->prog.vs = vc4_get_compiled_shader(vc4, QSTAGE_VERT, &key->base);
+ struct vc4_compiled_shader *vs =
+ vc4_get_compiled_shader(vc4, QSTAGE_VERT, &key->base);
+ if (vs != vc4->prog.vs) {
+ vc4->prog.vs = vs;
+ vc4->dirty |= VC4_DIRTY_COMPILED_VS;
+ }
+
key->is_coord = true;
- vc4->prog.cs = vc4_get_compiled_shader(vc4, QSTAGE_COORD, &key->base);
+ struct vc4_compiled_shader *cs =
+ vc4_get_compiled_shader(vc4, QSTAGE_COORD, &key->base);
+ if (cs != vc4->prog.cs) {
+ vc4->prog.cs = cs;
+ vc4->dirty |= VC4_DIRTY_COMPILED_CS;
+ }
}
void
vc4_update_compiled_vs(vc4, prim_mode);
}
-static unsigned
-fs_cache_hash(void *key)
+static uint32_t
+fs_cache_hash(const void *key)
{
return _mesa_hash_data(key, sizeof(struct vc4_fs_key));
}
-static unsigned
-vs_cache_hash(void *key)
+static uint32_t
+vs_cache_hash(const void *key)
{
return _mesa_hash_data(key, sizeof(struct vc4_vs_key));
}
-static int
-fs_cache_compare(void *key1, void *key2)
+static bool
+fs_cache_compare(const void *key1, const void *key2)
{
- return memcmp(key1, key2, sizeof(struct vc4_fs_key));
+ return memcmp(key1, key2, sizeof(struct vc4_fs_key)) == 0;
}
-static int
-vs_cache_compare(void *key1, void *key2)
+static bool
+vs_cache_compare(const void *key1, const void *key2)
{
- return memcmp(key1, key2, sizeof(struct vc4_vs_key));
+ return memcmp(key1, key2, sizeof(struct vc4_vs_key)) == 0;
}
-struct delete_state {
- struct vc4_context *vc4;
- struct vc4_uncompiled_shader *shader_state;
-};
-
-static enum pipe_error
-fs_delete_from_cache(void *in_key, void *in_value, void *data)
-{
- struct delete_state *del = data;
- struct vc4_fs_key *key = in_key;
- struct vc4_compiled_shader *shader = in_value;
-
- if (key->base.shader_state == data) {
- util_hash_table_remove(del->vc4->fs_cache, key);
- vc4_bo_unreference(&shader->bo);
- ralloc_free(shader);
- }
-
- return 0;
-}
-
-static enum pipe_error
-vs_delete_from_cache(void *in_key, void *in_value, void *data)
+static void
+delete_from_cache_if_matches(struct hash_table *ht,
+ struct hash_entry *entry,
+ struct vc4_uncompiled_shader *so)
{
- struct delete_state *del = data;
- struct vc4_vs_key *key = in_key;
- struct vc4_compiled_shader *shader = in_value;
+ const struct vc4_key *key = entry->key;
- if (key->base.shader_state == data) {
- util_hash_table_remove(del->vc4->vs_cache, key);
+ if (key->shader_state == so) {
+ struct vc4_compiled_shader *shader = entry->data;
+ _mesa_hash_table_remove(ht, entry);
vc4_bo_unreference(&shader->bo);
ralloc_free(shader);
}
-
- return 0;
}
static void
{
struct vc4_context *vc4 = vc4_context(pctx);
struct vc4_uncompiled_shader *so = hwcso;
- struct delete_state del;
- del.vc4 = vc4;
- del.shader_state = so;
- util_hash_table_foreach(vc4->fs_cache, fs_delete_from_cache, &del);
- util_hash_table_foreach(vc4->vs_cache, vs_delete_from_cache, &del);
+ struct hash_entry *entry;
+ hash_table_foreach(vc4->fs_cache, entry)
+ delete_from_cache_if_matches(vc4->fs_cache, entry, so);
+ hash_table_foreach(vc4->vs_cache, entry)
+ delete_from_cache_if_matches(vc4->vs_cache, entry, so);
if (so->twoside_tokens != so->base.tokens)
free((void *)so->twoside_tokens);
free(so);
}
-static uint32_t translate_wrap(uint32_t p_wrap, bool using_nearest)
-{
- switch (p_wrap) {
- case PIPE_TEX_WRAP_REPEAT:
- return 0;
- case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
- return 1;
- case PIPE_TEX_WRAP_MIRROR_REPEAT:
- return 2;
- case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
- return 3;
- case PIPE_TEX_WRAP_CLAMP:
- return (using_nearest ? 1 : 3);
- default:
- fprintf(stderr, "Unknown wrap mode %d\n", p_wrap);
- assert(!"not reached");
- return 0;
- }
-}
-
-static void
-write_texture_p0(struct vc4_context *vc4,
- struct vc4_texture_stateobj *texstate,
- uint32_t unit)
-{
- struct pipe_sampler_view *texture = texstate->textures[unit];
- struct vc4_resource *rsc = vc4_resource(texture->texture);
-
- cl_reloc(vc4, &vc4->uniforms, rsc->bo,
- VC4_SET_FIELD(rsc->slices[0].offset >> 12, VC4_TEX_P0_OFFSET) |
- VC4_SET_FIELD(texture->u.tex.last_level, VC4_TEX_P0_MIPLVLS) |
- VC4_SET_FIELD(texture->target == PIPE_TEXTURE_CUBE,
- VC4_TEX_P0_CMMODE) |
- VC4_SET_FIELD(rsc->vc4_format & 7, VC4_TEX_P0_TYPE));
-}
-
-static void
-write_texture_p1(struct vc4_context *vc4,
- struct vc4_texture_stateobj *texstate,
- uint32_t unit)
-{
- struct pipe_sampler_view *texture = texstate->textures[unit];
- struct vc4_resource *rsc = vc4_resource(texture->texture);
- struct pipe_sampler_state *sampler = texstate->samplers[unit];
- static const uint8_t minfilter_map[6] = {
- VC4_TEX_P1_MINFILT_NEAR_MIP_NEAR,
- VC4_TEX_P1_MINFILT_LIN_MIP_NEAR,
- VC4_TEX_P1_MINFILT_NEAR_MIP_LIN,
- VC4_TEX_P1_MINFILT_LIN_MIP_LIN,
- VC4_TEX_P1_MINFILT_NEAREST,
- VC4_TEX_P1_MINFILT_LINEAR,
- };
- static const uint32_t magfilter_map[] = {
- [PIPE_TEX_FILTER_NEAREST] = VC4_TEX_P1_MAGFILT_NEAREST,
- [PIPE_TEX_FILTER_LINEAR] = VC4_TEX_P1_MAGFILT_LINEAR,
- };
-
- bool either_nearest =
- (sampler->mag_img_filter == PIPE_TEX_MIPFILTER_NEAREST ||
- sampler->min_img_filter == PIPE_TEX_MIPFILTER_NEAREST);
-
- cl_u32(&vc4->uniforms,
- VC4_SET_FIELD(rsc->vc4_format >> 4, VC4_TEX_P1_TYPE4) |
- VC4_SET_FIELD(texture->texture->height0 & 2047,
- VC4_TEX_P1_HEIGHT) |
- VC4_SET_FIELD(texture->texture->width0 & 2047,
- VC4_TEX_P1_WIDTH) |
- VC4_SET_FIELD(magfilter_map[sampler->mag_img_filter],
- VC4_TEX_P1_MAGFILT) |
- VC4_SET_FIELD(minfilter_map[sampler->min_mip_filter * 2 +
- sampler->min_img_filter],
- VC4_TEX_P1_MINFILT) |
- VC4_SET_FIELD(translate_wrap(sampler->wrap_s, either_nearest),
- VC4_TEX_P1_WRAP_S) |
- VC4_SET_FIELD(translate_wrap(sampler->wrap_t, either_nearest),
- VC4_TEX_P1_WRAP_T));
-}
-
-static void
-write_texture_p2(struct vc4_context *vc4,
- struct vc4_texture_stateobj *texstate,
- uint32_t data)
-{
- uint32_t unit = data & 0xffff;
- struct pipe_sampler_view *texture = texstate->textures[unit];
- struct vc4_resource *rsc = vc4_resource(texture->texture);
-
- cl_u32(&vc4->uniforms,
- VC4_SET_FIELD(VC4_TEX_P2_PTYPE_CUBE_MAP_STRIDE,
- VC4_TEX_P2_PTYPE) |
- VC4_SET_FIELD(rsc->cube_map_stride >> 12, VC4_TEX_P2_CMST) |
- VC4_SET_FIELD((data >> 16) & 1, VC4_TEX_P2_BSLOD));
-}
-
-
-#define SWIZ(x,y,z,w) { \
- UTIL_FORMAT_SWIZZLE_##x, \
- UTIL_FORMAT_SWIZZLE_##y, \
- UTIL_FORMAT_SWIZZLE_##z, \
- UTIL_FORMAT_SWIZZLE_##w \
-}
-
-static void
-write_texture_border_color(struct vc4_context *vc4,
- struct vc4_texture_stateobj *texstate,
- uint32_t unit)
-{
- struct pipe_sampler_state *sampler = texstate->samplers[unit];
- struct pipe_sampler_view *texture = texstate->textures[unit];
- struct vc4_resource *rsc = vc4_resource(texture->texture);
- union util_color uc;
-
- const struct util_format_description *tex_format_desc =
- util_format_description(texture->format);
-
- float border_color[4];
- for (int i = 0; i < 4; i++)
- border_color[i] = sampler->border_color.f[i];
- if (util_format_is_srgb(texture->format)) {
- for (int i = 0; i < 3; i++)
- border_color[i] =
- util_format_linear_to_srgb_float(border_color[i]);
- }
-
- /* Turn the border color into the layout of channels that it would
- * have when stored as texture contents.
- */
- float storage_color[4];
- util_format_unswizzle_4f(storage_color,
- border_color,
- tex_format_desc->swizzle);
-
- /* Now, pack so that when the vc4_format-sampled texture contents are
- * replaced with our border color, the vc4_get_format_swizzle()
- * swizzling will get the right channels.
- */
- if (util_format_is_depth_or_stencil(texture->format)) {
- uc.ui[0] = util_pack_z(PIPE_FORMAT_Z24X8_UNORM,
- sampler->border_color.f[0]) << 8;
- } else {
- switch (rsc->vc4_format) {
- default:
- case VC4_TEXTURE_TYPE_RGBA8888:
- util_pack_color(storage_color,
- PIPE_FORMAT_R8G8B8A8_UNORM, &uc);
- break;
- case VC4_TEXTURE_TYPE_RGBA4444:
- util_pack_color(storage_color,
- PIPE_FORMAT_A8B8G8R8_UNORM, &uc);
- break;
- case VC4_TEXTURE_TYPE_RGB565:
- util_pack_color(storage_color,
- PIPE_FORMAT_B8G8R8A8_UNORM, &uc);
- break;
- case VC4_TEXTURE_TYPE_ALPHA:
- uc.ui[0] = float_to_ubyte(storage_color[0]) << 24;
- break;
- case VC4_TEXTURE_TYPE_LUMALPHA:
- uc.ui[0] = ((float_to_ubyte(storage_color[1]) << 24) |
- (float_to_ubyte(storage_color[0]) << 0));
- break;
- }
- }
-
- cl_u32(&vc4->uniforms, uc.ui[0]);
-}
-
-static uint32_t
-get_texrect_scale(struct vc4_texture_stateobj *texstate,
- enum quniform_contents contents,
- uint32_t data)
-{
- struct pipe_sampler_view *texture = texstate->textures[data];
- uint32_t dim;
-
- if (contents == QUNIFORM_TEXRECT_SCALE_X)
- dim = texture->texture->width0;
- else
- dim = texture->texture->height0;
-
- return fui(1.0f / dim);
-}
-
-void
-vc4_write_uniforms(struct vc4_context *vc4, struct vc4_compiled_shader *shader,
- struct vc4_constbuf_stateobj *cb,
- struct vc4_texture_stateobj *texstate)
-{
- struct vc4_shader_uniform_info *uinfo = &shader->uniforms;
- const uint32_t *gallium_uniforms = cb->cb[0].user_buffer;
-
- cl_start_shader_reloc(&vc4->uniforms, uinfo->num_texture_samples);
-
- for (int i = 0; i < uinfo->count; i++) {
-
- switch (uinfo->contents[i]) {
- case QUNIFORM_CONSTANT:
- cl_u32(&vc4->uniforms, uinfo->data[i]);
- break;
- case QUNIFORM_UNIFORM:
- cl_u32(&vc4->uniforms,
- gallium_uniforms[uinfo->data[i]]);
- break;
- case QUNIFORM_VIEWPORT_X_SCALE:
- cl_f(&vc4->uniforms, vc4->viewport.scale[0] * 16.0f);
- break;
- case QUNIFORM_VIEWPORT_Y_SCALE:
- cl_f(&vc4->uniforms, vc4->viewport.scale[1] * 16.0f);
- break;
-
- case QUNIFORM_VIEWPORT_Z_OFFSET:
- cl_f(&vc4->uniforms, vc4->viewport.translate[2]);
- break;
- case QUNIFORM_VIEWPORT_Z_SCALE:
- cl_f(&vc4->uniforms, vc4->viewport.scale[2]);
- break;
-
- case QUNIFORM_USER_CLIP_PLANE:
- cl_f(&vc4->uniforms,
- vc4->clip.ucp[uinfo->data[i] / 4][uinfo->data[i] % 4]);
- break;
-
- case QUNIFORM_TEXTURE_CONFIG_P0:
- write_texture_p0(vc4, texstate, uinfo->data[i]);
- break;
-
- case QUNIFORM_TEXTURE_CONFIG_P1:
- write_texture_p1(vc4, texstate, uinfo->data[i]);
- break;
-
- case QUNIFORM_TEXTURE_CONFIG_P2:
- write_texture_p2(vc4, texstate, uinfo->data[i]);
- break;
-
- case QUNIFORM_TEXTURE_BORDER_COLOR:
- write_texture_border_color(vc4, texstate, uinfo->data[i]);
- break;
-
- case QUNIFORM_TEXRECT_SCALE_X:
- case QUNIFORM_TEXRECT_SCALE_Y:
- cl_u32(&vc4->uniforms,
- get_texrect_scale(texstate,
- uinfo->contents[i],
- uinfo->data[i]));
- break;
-
- case QUNIFORM_BLEND_CONST_COLOR:
- cl_f(&vc4->uniforms,
- vc4->blend_color.color[uinfo->data[i]]);
- break;
-
- case QUNIFORM_STENCIL:
- cl_u32(&vc4->uniforms,
- vc4->zsa->stencil_uniforms[uinfo->data[i]] |
- (uinfo->data[i] <= 1 ?
- (vc4->stencil_ref.ref_value[uinfo->data[i]] << 8) :
- 0));
- break;
-
- case QUNIFORM_ALPHA_REF:
- cl_f(&vc4->uniforms, vc4->zsa->base.alpha.ref_value);
- break;
- }
-#if 0
- uint32_t written_val = *(uint32_t *)(vc4->uniforms.next - 4);
- fprintf(stderr, "%p: %d / 0x%08x (%f)\n",
- shader, i, written_val, uif(written_val));
-#endif
- }
-}
-
static void
vc4_fp_state_bind(struct pipe_context *pctx, void *hwcso)
{
struct vc4_context *vc4 = vc4_context(pctx);
vc4->prog.bind_fs = hwcso;
- vc4->prog.dirty |= VC4_SHADER_DIRTY_FP;
- vc4->dirty |= VC4_DIRTY_PROG;
+ vc4->dirty |= VC4_DIRTY_UNCOMPILED_FS;
}
static void
{
struct vc4_context *vc4 = vc4_context(pctx);
vc4->prog.bind_vs = hwcso;
- vc4->prog.dirty |= VC4_SHADER_DIRTY_VP;
- vc4->dirty |= VC4_DIRTY_PROG;
+ vc4->dirty |= VC4_DIRTY_UNCOMPILED_VS;
}
void
pctx->bind_fs_state = vc4_fp_state_bind;
pctx->bind_vs_state = vc4_vp_state_bind;
- vc4->fs_cache = util_hash_table_create(fs_cache_hash, fs_cache_compare);
- vc4->vs_cache = util_hash_table_create(vs_cache_hash, vs_cache_compare);
+ vc4->fs_cache = _mesa_hash_table_create(pctx, fs_cache_hash,
+ fs_cache_compare);
+ vc4->vs_cache = _mesa_hash_table_create(pctx, vs_cache_hash,
+ vs_cache_compare);
+}
+
+void
+vc4_program_fini(struct pipe_context *pctx)
+{
+ struct vc4_context *vc4 = vc4_context(pctx);
+
+ struct hash_entry *entry;
+ hash_table_foreach(vc4->fs_cache, entry) {
+ struct vc4_compiled_shader *shader = entry->data;
+ vc4_bo_unreference(&shader->bo);
+ ralloc_free(shader);
+ _mesa_hash_table_remove(vc4->fs_cache, entry);
+ }
+
+ hash_table_foreach(vc4->vs_cache, entry) {
+ struct vc4_compiled_shader *shader = entry->data;
+ vc4_bo_unreference(&shader->bo);
+ ralloc_free(shader);
+ _mesa_hash_table_remove(vc4->vs_cache, entry);
+ }
}
projects / mesa.git / blobdiff