*/
#include <inttypes.h>
-#include "pipe/p_state.h"
#include "util/u_format.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 "compiler/nir/nir.h"
+#include "compiler/nir/nir_builder.h"
#include "nir/tgsi_to_nir.h"
-
#include "vc4_context.h"
#include "vc4_qpu.h"
#include "vc4_qir.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;
- uint8_t logicop_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
-indirect_uniform_load(struct vc4_compile *c,
- struct qreg indirect_offset,
- unsigned offset)
+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->dst_offset = c->next_ubo_dst_offset;
c->next_ubo_dst_offset += range->size;
c->num_ubo_ranges++;
- };
+ }
offset -= range->src_offset;
- /* Translate the user's TGSI register index from the TGSI register
- * base to a byte offset.
- */
- indirect_offset = qir_SHL(c, indirect_offset, qir_uniform_ui(c, 4));
/* Adjust for where we stored the TGSI register base. */
indirect_offset = qir_ADD(c, indirect_offset,
qir_uniform_ui(c, (range->dst_offset +
offset)));
+
+ /* 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)));
qir_TEX_DIRECT(c, indirect_offset, qir_uniform(c, QUNIFORM_UBO_ADDR, 0));
- struct qreg r4 = qir_TEX_RESULT(c);
c->num_texture_samples++;
- return qir_MOV(c, r4);
+ return qir_TEX_RESULT(c);
}
-static struct qreg *
-ntq_get_dest(struct vc4_compile *c, nir_dest dest)
+nir_ssa_def *vc4_nir_get_state_uniform(struct nir_builder *b,
+ enum quniform_contents contents)
{
- assert(!dest.is_ssa);
- nir_register *reg = dest.reg.reg;
- struct hash_entry *entry = _mesa_hash_table_search(c->def_ht, reg);
- assert(reg->num_array_elems == 0);
- assert(dest.reg.base_offset == 0);
+ nir_intrinsic_instr *intr =
+ nir_intrinsic_instr_create(b->shader,
+ nir_intrinsic_load_uniform);
+ intr->const_index[0] = (VC4_NIR_STATE_UNIFORM_OFFSET + contents) * 4;
+ intr->num_components = 1;
+ intr->src[0] = nir_src_for_ssa(nir_imm_int(b, 0));
+ nir_ssa_dest_init(&intr->instr, &intr->dest, 1, 32, NULL);
+ nir_builder_instr_insert(b, &intr->instr);
+ return &intr->dest.ssa;
+}
- struct qreg *qregs = entry->data;
+nir_ssa_def *
+vc4_nir_get_swizzled_channel(nir_builder *b, nir_ssa_def **srcs, int swiz)
+{
+ switch (swiz) {
+ default:
+ case UTIL_FORMAT_SWIZZLE_NONE:
+ fprintf(stderr, "warning: unknown swizzle\n");
+ /* FALLTHROUGH */
+ case UTIL_FORMAT_SWIZZLE_0:
+ return nir_imm_float(b, 0.0);
+ case UTIL_FORMAT_SWIZZLE_1:
+ return nir_imm_float(b, 1.0);
+ case UTIL_FORMAT_SWIZZLE_X:
+ case UTIL_FORMAT_SWIZZLE_Y:
+ case UTIL_FORMAT_SWIZZLE_Z:
+ case UTIL_FORMAT_SWIZZLE_W:
+ return srcs[swiz];
+ }
+}
+
+static struct qreg *
+ntq_init_ssa_def(struct vc4_compile *c, nir_ssa_def *def)
+{
+ 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 *
+ntq_get_dest(struct vc4_compile *c, nir_dest *dest)
+{
+ 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
ntq_get_src(struct vc4_compile *c, nir_src src, int i)
{
return r;
};
-static struct qreg
-get_swizzled_channel(struct vc4_compile *c,
- struct qreg *srcs, int swiz)
-{
- switch (swiz) {
- default:
- case UTIL_FORMAT_SWIZZLE_NONE:
- fprintf(stderr, "warning: unknown swizzle\n");
- /* FALLTHROUGH */
- case UTIL_FORMAT_SWIZZLE_0:
- return qir_uniform_f(c, 0.0);
- case UTIL_FORMAT_SWIZZLE_1:
- return qir_uniform_f(c, 1.0);
- case UTIL_FORMAT_SWIZZLE_X:
- case UTIL_FORMAT_SWIZZLE_Y:
- case UTIL_FORMAT_SWIZZLE_Z:
- case UTIL_FORMAT_SWIZZLE_W:
- return srcs[swiz];
- }
-}
-
static inline struct qreg
qir_SAT(struct vc4_compile *c, struct qreg val)
{
qir_uniform_f(c, 2.4));
qir_SF(c, qir_FSUB(c, srgb, qir_uniform_f(c, 0.04045)));
- return qir_SEL_X_Y_NS(c, low, high);
-}
-
-static struct qreg
-qir_srgb_encode(struct vc4_compile *c, struct qreg linear)
-{
- struct qreg low = qir_FMUL(c, linear, qir_uniform_f(c, 12.92));
- struct qreg high = qir_FSUB(c,
- qir_FMUL(c,
- qir_uniform_f(c, 1.055),
- qir_POW(c,
- linear,
- qir_uniform_f(c, 0.41666))),
- qir_uniform_f(c, 0.055));
-
- qir_SF(c, qir_FSUB(c, linear, qir_uniform_f(c, 0.0031308)));
- return qir_SEL_X_Y_NS(c, low, high);
+ return qir_SEL(c, QPU_COND_NS, low, high);
}
static struct qreg
qir_uniform_ui(c, 24)));
}
+static struct qreg
+ntq_scale_depth_texture(struct vc4_compile *c, struct qreg src)
+{
+ struct qreg depthf = qir_ITOF(c, qir_SHR(c, src,
+ qir_uniform_ui(c, 8)));
+ return qir_FMUL(c, depthf, qir_uniform_f(c, 1.0f/0xffffff));
+}
+
+/**
+ * Emits a lowered TXF_MS from an MSAA texture.
+ *
+ * The addressing math has been lowered in NIR, and now we just need to read
+ * it like a UBO.
+ */
+static void
+ntq_emit_txf(struct vc4_compile *c, nir_tex_instr *instr)
+{
+ uint32_t tile_width = 32;
+ uint32_t tile_height = 32;
+ uint32_t tile_size = (tile_height * tile_width *
+ VC4_MAX_SAMPLES * sizeof(uint32_t));
+
+ unsigned unit = instr->texture_index;
+ uint32_t w = align(c->key->tex[unit].msaa_width, tile_width);
+ uint32_t w_tiles = w / tile_width;
+ uint32_t h = align(c->key->tex[unit].msaa_height, tile_height);
+ uint32_t h_tiles = h / tile_height;
+ uint32_t size = w_tiles * h_tiles * tile_size;
+
+ struct qreg addr;
+ assert(instr->num_srcs == 1);
+ assert(instr->src[0].src_type == nir_tex_src_coord);
+ addr = ntq_get_src(c, instr->src[0].src, 0);
+
+ /* Perform the clamping required by kernel validation. */
+ addr = qir_MAX(c, addr, qir_uniform_ui(c, 0));
+ addr = qir_MIN(c, addr, qir_uniform_ui(c, size - 4));
+
+ qir_TEX_DIRECT(c, addr, qir_uniform(c, QUNIFORM_TEXTURE_MSAA_ADDR, unit));
+
+ struct qreg tex = qir_TEX_RESULT(c);
+ c->num_texture_samples++;
+
+ struct qreg *dest = ntq_get_dest(c, &instr->dest);
+ enum pipe_format format = c->key->tex[unit].format;
+ if (util_format_is_depth_or_stencil(format)) {
+ struct qreg scaled = ntq_scale_depth_texture(c, tex);
+ for (int i = 0; i < 4; i++)
+ dest[i] = scaled;
+ } else {
+ for (int i = 0; i < 4; i++)
+ dest[i] = qir_UNPACK_8_F(c, tex, i);
+ }
+
+ for (int i = 0; i < 4; i++) {
+ if (c->tex_srgb_decode[unit] & (1 << i))
+ dest[i] = qir_srgb_decode(c, dest[i]);
+ }
+}
+
static void
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;
+ unsigned unit = instr->texture_index;
+
+ if (instr->op == nir_texop_txf) {
+ ntq_emit_txf(c, instr);
+ return;
+ }
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)
+ 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);
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];
+ struct qreg *dest = ntq_get_dest(c, &instr->dest);
if (util_format_is_depth_or_stencil(format)) {
- struct qreg depthf = qir_ITOF(c, qir_SHR(c, r4,
- qir_uniform_ui(c, 8)));
- struct qreg normalized = qir_FMUL(c, depthf,
- qir_uniform_f(c, 1.0f/0xffffff));
-
+ struct qreg normalized = ntq_scale_depth_texture(c, tex);
struct qreg depth_output;
- struct qreg one = qir_uniform_f(c, 1.0f);
+ struct qreg u0 = qir_uniform_f(c, 0.0f);
+ struct qreg u1 = qir_uniform_f(c, 1.0f);
if (c->key->tex[unit].compare_mode) {
if (has_proj)
compare = qir_FMUL(c, compare, proj);
depth_output = qir_uniform_f(c, 0.0f);
break;
case PIPE_FUNC_ALWAYS:
- depth_output = one;
+ depth_output = u1;
break;
case PIPE_FUNC_EQUAL:
qir_SF(c, qir_FSUB(c, compare, normalized));
- depth_output = qir_SEL_X_0_ZS(c, one);
+ depth_output = qir_SEL(c, QPU_COND_ZS, u1, u0);
break;
case PIPE_FUNC_NOTEQUAL:
qir_SF(c, qir_FSUB(c, compare, normalized));
- depth_output = qir_SEL_X_0_ZC(c, one);
+ depth_output = qir_SEL(c, QPU_COND_ZC, u1, u0);
break;
case PIPE_FUNC_GREATER:
qir_SF(c, qir_FSUB(c, compare, normalized));
- depth_output = qir_SEL_X_0_NC(c, one);
+ depth_output = qir_SEL(c, QPU_COND_NC, u1, u0);
break;
case PIPE_FUNC_GEQUAL:
qir_SF(c, qir_FSUB(c, normalized, compare));
- depth_output = qir_SEL_X_0_NS(c, one);
+ depth_output = qir_SEL(c, QPU_COND_NS, u1, u0);
break;
case PIPE_FUNC_LESS:
qir_SF(c, qir_FSUB(c, compare, normalized));
- depth_output = qir_SEL_X_0_NS(c, one);
+ depth_output = qir_SEL(c, QPU_COND_NS, u1, u0);
break;
case PIPE_FUNC_LEQUAL:
qir_SF(c, qir_FSUB(c, normalized, compare));
- depth_output = qir_SEL_X_0_NC(c, one);
+ depth_output = qir_SEL(c, QPU_COND_NC, u1, u0);
break;
}
} else {
}
for (int i = 0; i < 4; i++)
- unpacked[i] = depth_output;
+ dest[i] = depth_output;
} else {
for (int i = 0; i < 4; i++)
- unpacked[i] = qir_R4_UNPACK(c, r4, i);
+ dest[i] = qir_UNPACK_8_F(c, tex, i);
}
- const uint8_t *format_swiz = vc4_get_format_swizzle(format);
- struct qreg texture_output[4];
for (int i = 0; i < 4; i++) {
- texture_output[i] = get_swizzled_channel(c, unpacked,
- format_swiz[i]);
- }
-
- if (util_format_is_srgb(format)) {
- for (int i = 0; i < 3; i++)
- texture_output[i] = qir_srgb_decode(c,
- texture_output[i]);
- }
-
- struct qreg *dest = ntq_get_dest(c, instr->dest);
- for (int i = 0; i < 4; i++) {
- dest[i] = get_swizzled_channel(c, texture_output,
- c->key->tex[unit].swizzle[i]);
+ if (c->tex_srgb_decode[unit] & (1 << i))
+ dest[i] = qir_srgb_decode(c, dest[i]);
}
}
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)),
- diff);
+ return qir_SEL(c, QPU_COND_NS,
+ qir_FADD(c, diff, qir_uniform_f(c, 1.0)), diff);
}
/**
*/
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)),
- trunc);
+ return qir_SEL(c, QPU_COND_NS,
+ qir_FSUB(c, trunc, qir_uniform_f(c, 1.0)), trunc);
}
/**
*/
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)),
- trunc);
+ return qir_SEL(c, QPU_COND_NS,
+ qir_FADD(c, trunc, qir_uniform_f(c, 1.0)), trunc);
}
static struct qreg
struct qreg scaled_x =
qir_FMUL(c,
src,
- qir_uniform_f(c, 1.0f / (M_PI * 2.0f)));
+ qir_uniform_f(c, 1.0 / (M_PI * 2.0)));
struct qreg x = qir_FADD(c,
ntq_ffract(c, scaled_x),
static struct qreg
ntq_fsign(struct vc4_compile *c, struct qreg src)
{
- 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];
+ struct qreg t = qir_get_temp(c);
- /* 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;
- }
+ qir_SF(c, src);
+ qir_MOV_dest(c, t, qir_uniform_f(c, 0.0));
+ qir_MOV_dest(c, t, qir_uniform_f(c, 1.0))->cond = QPU_COND_ZC;
+ qir_MOV_dest(c, t, qir_uniform_f(c, -1.0))->cond = QPU_COND_NS;
+ return t;
}
static void
{
enum pipe_format format = c->vs_key->attr_formats[attr];
uint32_t attr_size = util_format_get_blocksize(format);
- struct qreg vpm_reads[4];
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->inputs[attr * 4 + i] = qir_MOV(c, vpm);
c->num_inputs++;
}
-
- bool format_warned = false;
- const struct util_format_description *desc =
- util_format_description(format);
-
- for (int i = 0; i < 4; i++) {
- uint8_t swiz = desc->swizzle[i];
- 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",
- attr, util_format_name(format));
- format_warned = true;
- }
- result = qir_uniform_f(c, 0.0);
- }
- c->inputs[attr * 4 + i] = result;
- }
}
static void
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)
+emit_fragment_varying(struct vc4_compile *c, gl_varying_slot slot,
+ uint8_t swizzle)
{
- uint32_t i = c->num_input_semantics++;
+ uint32_t i = c->num_input_slots++;
struct qreg vary = {
QFILE_VARY,
i
};
- if (c->num_input_semantics >= c->input_semantics_array_size) {
- c->input_semantics_array_size =
- MAX2(4, c->input_semantics_array_size * 2);
+ if (c->num_input_slots >= c->input_slots_array_size) {
+ c->input_slots_array_size =
+ MAX2(4, c->input_slots_array_size * 2);
- c->input_semantics = reralloc(c, c->input_semantics,
- struct vc4_varying_semantic,
- c->input_semantics_array_size);
+ c->input_slots = reralloc(c, c->input_slots,
+ struct vc4_varying_slot,
+ c->input_slots_array_size);
}
- c->input_semantics[i].semantic = semantic;
- c->input_semantics[i].index = index;
- c->input_semantics[i].swizzle = swizzle;
+ c->input_slots[i].slot = slot;
+ c->input_slots[i].swizzle = swizzle;
return qir_VARY_ADD_C(c, qir_FMUL(c, vary, qir_FRAG_W(c)));
}
static void
-emit_fragment_input(struct vc4_compile *c, int attr,
- unsigned semantic_name, unsigned semantic_index)
+emit_fragment_input(struct vc4_compile *c, int attr, gl_varying_slot slot)
{
for (int i = 0; i < 4; i++) {
c->inputs[attr * 4 + i] =
- emit_fragment_varying(c,
- semantic_name,
- semantic_index,
- i);
+ emit_fragment_varying(c, slot, 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,
- uint8_t semantic_name,
- uint8_t semantic_index,
- uint8_t semantic_swizzle)
+ uint8_t slot,
+ uint8_t swizzle)
{
uint32_t old_array_size = c->outputs_array_size;
resize_qreg_array(c, &c->outputs, &c->outputs_array_size,
decl_offset + 1);
if (old_array_size != c->outputs_array_size) {
- c->output_semantics = reralloc(c,
- c->output_semantics,
- struct vc4_varying_semantic,
- c->outputs_array_size);
+ c->output_slots = reralloc(c,
+ c->output_slots,
+ struct vc4_varying_slot,
+ c->outputs_array_size);
}
- c->output_semantics[decl_offset].semantic = semantic_name;
- c->output_semantics[decl_offset].index = semantic_index;
- c->output_semantics[decl_offset].swizzle = semantic_swizzle;
+ c->output_slots[decl_offset].slot = slot;
+ c->output_slots[decl_offset].swizzle = swizzle;
}
static void
c->ubo_ranges[array_id].used = false;
}
+static bool
+ntq_src_is_only_ssa_def_user(nir_src *src)
+{
+ if (!src->is_ssa)
+ return false;
+
+ if (!list_empty(&src->ssa->if_uses))
+ return false;
+
+ return (src->ssa->uses.next == &src->use_link &&
+ src->ssa->uses.next->next == &src->ssa->uses);
+}
+
+/**
+ * In general, emits a nir_pack_unorm_4x8 as a series of MOVs with the pack
+ * bit set.
+ *
+ * However, as an optimization, it tries to find the instructions generating
+ * the sources to be packed and just emit the pack flag there, if possible.
+ */
+static void
+ntq_emit_pack_unorm_4x8(struct vc4_compile *c, nir_alu_instr *instr)
+{
+ struct qreg result = qir_get_temp(c);
+ struct nir_alu_instr *vec4 = NULL;
+
+ /* If packing from a vec4 op (as expected), identify it so that we can
+ * peek back at what generated its sources.
+ */
+ if (instr->src[0].src.is_ssa &&
+ instr->src[0].src.ssa->parent_instr->type == nir_instr_type_alu &&
+ nir_instr_as_alu(instr->src[0].src.ssa->parent_instr)->op ==
+ nir_op_vec4) {
+ vec4 = nir_instr_as_alu(instr->src[0].src.ssa->parent_instr);
+ }
+
+ /* If the pack is replicating the same channel 4 times, use the 8888
+ * pack flag. This is common for blending using the alpha
+ * channel.
+ */
+ if (instr->src[0].swizzle[0] == instr->src[0].swizzle[1] &&
+ instr->src[0].swizzle[0] == instr->src[0].swizzle[2] &&
+ instr->src[0].swizzle[0] == instr->src[0].swizzle[3]) {
+ struct qreg *dest = ntq_get_dest(c, &instr->dest.dest);
+ *dest = qir_PACK_8888_F(c,
+ ntq_get_src(c, instr->src[0].src,
+ instr->src[0].swizzle[0]));
+ return;
+ }
+
+ for (int i = 0; i < 4; i++) {
+ int swiz = instr->src[0].swizzle[i];
+ struct qreg src;
+ if (vec4) {
+ src = ntq_get_src(c, vec4->src[swiz].src,
+ vec4->src[swiz].swizzle[0]);
+ } else {
+ src = ntq_get_src(c, instr->src[0].src, swiz);
+ }
+
+ if (vec4 &&
+ ntq_src_is_only_ssa_def_user(&vec4->src[swiz].src) &&
+ src.file == QFILE_TEMP &&
+ c->defs[src.index] &&
+ qir_is_mul(c->defs[src.index]) &&
+ !c->defs[src.index]->dst.pack) {
+ struct qinst *rewrite = c->defs[src.index];
+ c->defs[src.index] = NULL;
+ rewrite->dst = result;
+ rewrite->dst.pack = QPU_PACK_MUL_8A + i;
+ continue;
+ }
+
+ qir_PACK_8_F(c, result, src, i);
+ }
+
+ struct qreg *dest = ntq_get_dest(c, &instr->dest.dest);
+ *dest = result;
+}
+
+/** Handles sign-extended bitfield extracts for 16 bits. */
+static struct qreg
+ntq_emit_ibfe(struct vc4_compile *c, struct qreg base, struct qreg offset,
+ struct qreg bits)
+{
+ assert(bits.file == QFILE_UNIF &&
+ c->uniform_contents[bits.index] == QUNIFORM_CONSTANT &&
+ c->uniform_data[bits.index] == 16);
+
+ assert(offset.file == QFILE_UNIF &&
+ c->uniform_contents[offset.index] == QUNIFORM_CONSTANT);
+ int offset_bit = c->uniform_data[offset.index];
+ assert(offset_bit % 16 == 0);
+
+ return qir_UNPACK_16_I(c, base, offset_bit / 16);
+}
+
+/** Handles unsigned bitfield extracts for 8 bits. */
+static struct qreg
+ntq_emit_ubfe(struct vc4_compile *c, struct qreg base, struct qreg offset,
+ struct qreg bits)
+{
+ assert(bits.file == QFILE_UNIF &&
+ c->uniform_contents[bits.index] == QUNIFORM_CONSTANT &&
+ c->uniform_data[bits.index] == 8);
+
+ assert(offset.file == QFILE_UNIF &&
+ c->uniform_contents[offset.index] == QUNIFORM_CONSTANT);
+ int offset_bit = c->uniform_data[offset.index];
+ assert(offset_bit % 8 == 0);
+
+ return qir_UNPACK_8_I(c, base, offset_bit / 8);
+}
+
+/**
+ * If compare_instr is a valid comparison instruction, emits the
+ * compare_instr's comparison and returns the sel_instr's return value based
+ * on the compare_instr's result.
+ */
+static bool
+ntq_emit_comparison(struct vc4_compile *c, struct qreg *dest,
+ nir_alu_instr *compare_instr,
+ nir_alu_instr *sel_instr)
+{
+ enum qpu_cond cond;
+
+ switch (compare_instr->op) {
+ case nir_op_feq:
+ case nir_op_ieq:
+ case nir_op_seq:
+ cond = QPU_COND_ZS;
+ break;
+ case nir_op_fne:
+ case nir_op_ine:
+ case nir_op_sne:
+ cond = QPU_COND_ZC;
+ break;
+ case nir_op_fge:
+ case nir_op_ige:
+ case nir_op_uge:
+ case nir_op_sge:
+ cond = QPU_COND_NC;
+ break;
+ case nir_op_flt:
+ case nir_op_ilt:
+ case nir_op_slt:
+ cond = QPU_COND_NS;
+ break;
+ default:
+ return false;
+ }
+
+ struct qreg src0 = ntq_get_alu_src(c, compare_instr, 0);
+ struct qreg src1 = ntq_get_alu_src(c, compare_instr, 1);
+
+ unsigned unsized_type =
+ nir_alu_type_get_base_type(nir_op_infos[compare_instr->op].input_types[0]);
+ if (unsized_type == nir_type_float)
+ qir_SF(c, qir_FSUB(c, src0, src1));
+ else
+ qir_SF(c, qir_SUB(c, src0, src1));
+
+ switch (sel_instr->op) {
+ case nir_op_seq:
+ case nir_op_sne:
+ case nir_op_sge:
+ case nir_op_slt:
+ *dest = qir_SEL(c, cond,
+ qir_uniform_f(c, 1.0), qir_uniform_f(c, 0.0));
+ break;
+
+ case nir_op_bcsel:
+ *dest = qir_SEL(c, cond,
+ ntq_get_alu_src(c, sel_instr, 1),
+ ntq_get_alu_src(c, sel_instr, 2));
+ break;
+
+ default:
+ *dest = qir_SEL(c, cond,
+ qir_uniform_ui(c, ~0), qir_uniform_ui(c, 0));
+ break;
+ }
+
+ return true;
+}
+
+/**
+ * Attempts to fold a comparison generating a boolean result into the
+ * condition code for selecting between two values, instead of comparing the
+ * boolean result against 0 to generate the condition code.
+ */
+static struct qreg ntq_emit_bcsel(struct vc4_compile *c, nir_alu_instr *instr,
+ struct qreg *src)
+{
+ if (!instr->src[0].src.is_ssa)
+ goto out;
+ nir_alu_instr *compare =
+ nir_instr_as_alu(instr->src[0].src.ssa->parent_instr);
+ if (!compare)
+ goto out;
+
+ struct qreg dest;
+ if (ntq_emit_comparison(c, &dest, compare, instr))
+ return dest;
+
+out:
+ qir_SF(c, src[0]);
+ return qir_SEL(c, QPU_COND_NS, src[1], src[2]);
+}
+
static void
ntq_emit_alu(struct vc4_compile *c, nir_alu_instr *instr)
{
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);
+ 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;
}
+ if (instr->op == nir_op_pack_unorm_4x8) {
+ ntq_emit_pack_unorm_4x8(c, instr);
+ return;
+ }
+
+ if (instr->op == nir_op_unpack_unorm_4x8) {
+ struct qreg src = ntq_get_src(c, instr->src[0].src,
+ instr->src[0].swizzle[0]);
+ struct qreg *dest = ntq_get_dest(c, &instr->dest.dest);
+ for (int i = 0; i < 4; i++) {
+ if (instr->dest.write_mask & (1 << i))
+ dest[i] = qir_UNPACK_8_F(c, src, i);
+ }
+ return;
+ }
+
/* 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++) {
/* Pick the channel to store the output in. */
assert(!instr->dest.saturate);
- struct qreg *dest = ntq_get_dest(c, instr->dest.dest);
+ 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;
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));
+ *dest = qir_SEL(c, QPU_COND_ZC,
+ qir_uniform_ui(c, ~0),
+ qir_uniform_ui(c, 0));
break;
case nir_op_iadd:
break;
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_uge:
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));
+ if (!ntq_emit_comparison(c, dest, instr, instr)) {
+ fprintf(stderr, "Bad comparison instruction\n");
+ }
break;
case nir_op_bcsel:
- qir_SF(c, src[0]);
- *dest = qir_SEL_X_Y_NS(c, src[1], src[2]);
+ *dest = ntq_emit_bcsel(c, instr, src);
break;
case nir_op_fcsel:
qir_SF(c, src[0]);
- *dest = qir_SEL_X_Y_ZC(c, src[1], src[2]);
+ *dest = qir_SEL(c, QPU_COND_ZC, src[1], src[2]);
break;
case nir_op_frcp:
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();
- }
-}
-
-static struct qreg
-vc4_blend_channel(struct vc4_compile *c,
- struct qreg *dst,
- struct qreg *src,
- struct qreg val,
- unsigned factor,
- int channel)
-{
- switch(factor) {
- case PIPE_BLENDFACTOR_ONE:
- return val;
- case PIPE_BLENDFACTOR_SRC_COLOR:
- return qir_FMUL(c, val, src[channel]);
- case PIPE_BLENDFACTOR_SRC_ALPHA:
- return qir_FMUL(c, val, src[3]);
- case PIPE_BLENDFACTOR_DST_ALPHA:
- return qir_FMUL(c, val, dst[3]);
- case PIPE_BLENDFACTOR_DST_COLOR:
- return qir_FMUL(c, val, dst[channel]);
- case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
- 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,
- qir_uniform(c, QUNIFORM_BLEND_CONST_COLOR,
- channel));
- case PIPE_BLENDFACTOR_CONST_ALPHA:
- return qir_FMUL(c, val,
- 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:
- return qir_FMUL(c, val, qir_FSUB(c, qir_uniform_f(c, 1.0),
- src[channel]));
- case PIPE_BLENDFACTOR_INV_SRC_ALPHA:
- return qir_FMUL(c, val, qir_FSUB(c, qir_uniform_f(c, 1.0),
- src[3]));
- case PIPE_BLENDFACTOR_INV_DST_ALPHA:
- return qir_FMUL(c, val, qir_FSUB(c, qir_uniform_f(c, 1.0),
- dst[3]));
- case PIPE_BLENDFACTOR_INV_DST_COLOR:
- return qir_FMUL(c, val, qir_FSUB(c, qir_uniform_f(c, 1.0),
- dst[channel]));
- case PIPE_BLENDFACTOR_INV_CONST_COLOR:
- return qir_FMUL(c, val,
- qir_FSUB(c, qir_uniform_f(c, 1.0),
- 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),
- qir_uniform(c,
- QUNIFORM_BLEND_CONST_COLOR,
- 3)));
-
- default:
- case PIPE_BLENDFACTOR_SRC1_COLOR:
- case PIPE_BLENDFACTOR_SRC1_ALPHA:
- case PIPE_BLENDFACTOR_INV_SRC1_COLOR:
- case PIPE_BLENDFACTOR_INV_SRC1_ALPHA:
- /* Unsupported. */
- fprintf(stderr, "Unknown blend factor %d\n", factor);
- return val;
- }
-}
-
-static struct qreg
-vc4_blend_func(struct vc4_compile *c,
- struct qreg src, struct qreg dst,
- unsigned func)
-{
- switch (func) {
- case PIPE_BLEND_ADD:
- return qir_FADD(c, src, dst);
- case PIPE_BLEND_SUBTRACT:
- return qir_FSUB(c, src, dst);
- case PIPE_BLEND_REVERSE_SUBTRACT:
- return qir_FSUB(c, dst, src);
- case PIPE_BLEND_MIN:
- return qir_FMIN(c, src, dst);
- case PIPE_BLEND_MAX:
- return qir_FMAX(c, src, dst);
-
- default:
- /* Unsupported. */
- fprintf(stderr, "Unknown blend func %d\n", func);
- return src;
-
- }
-}
-
-/**
- * Implements fixed function blending in shader code.
- *
- * VC4 doesn't have any hardware support for blending. Instead, you read the
- * current contents of the destination from the tile buffer after having
- * waited for the scoreboard (which is handled by vc4_qpu_emit.c), then do
- * math using your output color and that destination value, and update the
- * output color appropriately.
- */
-static void
-vc4_blend(struct vc4_compile *c, struct qreg *result,
- struct qreg *dst_color, struct qreg *src_color)
-{
- struct pipe_rt_blend_state *blend = &c->fs_key->blend;
-
- if (!blend->blend_enable) {
- for (int i = 0; i < 4; i++)
- result[i] = src_color[i];
- return;
- }
-
- struct qreg clamped_src[4];
- struct qreg clamped_dst[4];
- for (int i = 0; i < 4; i++) {
- clamped_src[i] = qir_SAT(c, src_color[i]);
- clamped_dst[i] = qir_SAT(c, dst_color[i]);
- }
- src_color = clamped_src;
- dst_color = clamped_dst;
-
- struct qreg src_blend[4], dst_blend[4];
- for (int i = 0; i < 3; i++) {
- src_blend[i] = vc4_blend_channel(c,
- dst_color, src_color,
- src_color[i],
- blend->rgb_src_factor, i);
- dst_blend[i] = vc4_blend_channel(c,
- dst_color, src_color,
- dst_color[i],
- blend->rgb_dst_factor, i);
- }
- src_blend[3] = vc4_blend_channel(c,
- dst_color, src_color,
- src_color[3],
- blend->alpha_src_factor, 3);
- dst_blend[3] = vc4_blend_channel(c,
- dst_color, src_color,
- dst_color[3],
- blend->alpha_dst_factor, 3);
-
- for (int i = 0; i < 3; i++) {
- result[i] = vc4_blend_func(c,
- src_blend[i], dst_blend[i],
- blend->rgb_func);
- }
- result[3] = vc4_blend_func(c,
- src_blend[3], dst_blend[3],
- blend->alpha_func);
-}
-
-static void
-clip_distance_discard(struct vc4_compile *c)
-{
- for (int i = 0; i < PIPE_MAX_CLIP_PLANES; i++) {
- if (!(c->key->ucp_enables & (1 << i)))
- continue;
-
- struct qreg dist = emit_fragment_varying(c,
- TGSI_SEMANTIC_CLIPDIST,
- i,
- TGSI_SWIZZLE_X);
-
- qir_SF(c, dist);
-
- if (c->discard.file == QFILE_NULL)
- c->discard = qir_uniform_ui(c, 0);
-
- c->discard = qir_SEL_X_Y_NS(c, qir_uniform_ui(c, ~0),
- c->discard);
- }
-}
-
-static void
-alpha_test_discard(struct vc4_compile *c)
-{
- struct qreg src_alpha;
- struct qreg alpha_ref = qir_uniform(c, QUNIFORM_ALPHA_REF, 0);
-
- if (!c->fs_key->alpha_test)
- return;
-
- if (c->output_color_index != -1)
- src_alpha = c->outputs[c->output_color_index + 3];
- else
- src_alpha = qir_uniform_f(c, 1.0);
-
- if (c->discard.file == QFILE_NULL)
- c->discard = qir_uniform_ui(c, 0);
-
- switch (c->fs_key->alpha_test_func) {
- case PIPE_FUNC_NEVER:
- c->discard = qir_uniform_ui(c, ~0);
- break;
- case PIPE_FUNC_ALWAYS:
+ case nir_op_ibitfield_extract:
+ *dest = ntq_emit_ibfe(c, src[0], src[1], src[2]);
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_ui(c, ~0));
+
+ case nir_op_ubitfield_extract:
+ *dest = ntq_emit_ubfe(c, src[0], src[1], src[2]);
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_ui(c, ~0));
+
+ case nir_op_usadd_4x8:
+ *dest = qir_V8ADDS(c, src[0], src[1]);
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_ui(c, ~0));
+
+ case nir_op_ussub_4x8:
+ *dest = qir_V8SUBS(c, src[0], src[1]);
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_ui(c, ~0));
+
+ case nir_op_umin_4x8:
+ *dest = qir_V8MIN(c, src[0], src[1]);
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_ui(c, ~0));
+
+ case nir_op_umax_4x8:
+ *dest = qir_V8MAX(c, src[0], src[1]);
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_ui(c, ~0));
+
+ case nir_op_umul_unorm_4x8:
+ *dest = qir_V8MULD(c, src[0], src[1]);
break;
- }
-}
-static struct qreg
-vc4_logicop(struct vc4_compile *c, struct qreg src, struct qreg dst)
-{
- 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;
+ fprintf(stderr, "unknown NIR ALU inst: ");
+ nir_print_instr(&instr->instr, stderr);
+ fprintf(stderr, "\n");
+ abort();
}
}
static void
emit_frag_end(struct vc4_compile *c)
{
- clip_distance_discard(c);
- alpha_test_discard(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 ||
- c->fs_key->logicop_func != PIPE_LOGICOP_COPY) {
- struct qreg r4 = qir_TLB_COLOR_READ(c);
- for (int i = 0; i < 4; i++)
- tlb_read_color[i] = qir_R4_UNPACK(c, r4, i);
- for (int i = 0; i < 4; i++) {
- dst_color[i] = get_swizzled_channel(c,
- tlb_read_color,
- format_swiz[i]);
- if (util_format_is_srgb(color_format) && i != 3) {
- linear_dst_color[i] =
- qir_srgb_decode(c, dst_color[i]);
- } else {
- linear_dst_color[i] = dst_color[i];
- }
- }
-
- /* Save the packed value for logic ops. Can't reuse r4
- * because other things might smash it (like sRGB)
- */
- packed_dst_color = qir_MOV(c, r4);
- }
-
- 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));
-
- if (util_format_is_srgb(color_format)) {
- for (int i = 0; i < 3; i++)
- blend_color[i] = qir_srgb_encode(c, blend_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.
- */
- if (0) {
- blend_color[0] = qir_uniform_f(c, 1.0);
- blend_color[1] = qir_uniform_f(c, 0.0);
- blend_color[2] = qir_uniform_f(c, 1.0);
- blend_color[3] = qir_uniform_f(c, 0.5);
+ struct qreg color;
+ if (c->output_color_index != -1) {
+ color = c->outputs[c->output_color_index];
+ } else {
+ color = qir_uniform_ui(c, 0);
}
- struct qreg swizzled_outputs[4];
- for (int i = 0; i < 4; i++) {
- swizzled_outputs[i] = get_swizzled_channel(c, blend_color,
- format_swiz[i]);
+ uint32_t discard_cond = QPU_COND_ALWAYS;
+ if (c->discard.file != QFILE_NULL) {
+ qir_SF(c, c->discard);
+ discard_cond = QPU_COND_ZS;
}
- if (c->discard.file != QFILE_NULL)
- qir_TLB_DISCARD_SETUP(c, c->discard);
-
if (c->fs_key->stencil_enabled) {
qir_TLB_STENCIL_SETUP(c, qir_uniform(c, QUNIFORM_STENCIL, 0));
if (c->fs_key->stencil_twoside) {
}
}
+ if (c->output_sample_mask_index != -1) {
+ qir_MS_MASK(c, c->outputs[c->output_sample_mask_index]);
+ }
+
if (c->fs_key->depth_enabled) {
struct qreg z;
if (c->output_position_index != -1) {
} else {
z = qir_FRAG_Z(c);
}
- qir_TLB_Z_WRITE(c, z);
- }
-
- 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);
- }
+ struct qinst *inst = qir_TLB_Z_WRITE(c, z);
+ inst->cond = discard_cond;
}
- 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 (!c->msaa_per_sample_output) {
+ struct qinst *inst = qir_TLB_COLOR_WRITE(c, color);
+ inst->cond = discard_cond;
+ } else {
+ for (int i = 0; i < VC4_MAX_SAMPLES; i++) {
+ struct qinst *inst = qir_TLB_COLOR_WRITE_MS(c, c->sample_colors[i]);
+ inst->cond = discard_cond;
}
}
- 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)));
- }
-
- qir_emit(c, qir_inst(QOP_TLB_COLOR_WRITE, c->undef,
- packed_color, c->undef));
}
static void
emit_scaled_viewport_write(struct vc4_compile *c, struct qreg rcp_w)
{
- struct qreg xyi[2];
+ struct qreg packed = qir_get_temp(c);
for (int i = 0; i < 2; i++) {
struct qreg scale =
qir_uniform(c, QUNIFORM_VIEWPORT_X_SCALE + i, 0);
- xyi[i] = qir_FTOI(c, qir_FMUL(c,
- qir_FMUL(c,
- c->outputs[c->output_position_index + i],
- scale),
- rcp_w));
+ struct qreg packed_chan = packed;
+ packed_chan.pack = QPU_PACK_A_16A + i;
+
+ qir_FTOI_dest(c, packed_chan,
+ qir_FMUL(c,
+ qir_FMUL(c,
+ c->outputs[c->output_position_index + i],
+ scale),
+ rcp_w));
}
- qir_VPM_WRITE(c, qir_PACK_SCALED(c, xyi[0], xyi[1]));
+ qir_VPM_WRITE(c, packed);
}
static void
struct qreg point_size;
if (c->output_point_size_index != -1)
- point_size = c->outputs[c->output_point_size_index + 3];
+ point_size = c->outputs[c->output_point_size_index];
else
point_size = qir_uniform_f(c, 1.0);
c->num_inputs++;
}
-static void
-emit_ucp_clipdistance(struct vc4_compile *c)
-{
- unsigned cv;
- if (c->output_clipvertex_index != -1)
- cv = c->output_clipvertex_index;
- else if (c->output_position_index != -1)
- cv = c->output_position_index;
- else
- return;
-
- for (int plane = 0; plane < PIPE_MAX_CLIP_PLANES; plane++) {
- if (!(c->key->ucp_enables & (1 << plane)))
- continue;
-
- /* Pick the next outputs[] that hasn't been written to, since
- * there are no other program writes left to be processed at
- * this point. If something had been declared but not written
- * (like a w component), we'll just smash over the top of it.
- */
- uint32_t output_index = c->num_outputs++;
- add_output(c, output_index,
- TGSI_SEMANTIC_CLIPDIST,
- 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 =
- qir_uniform(c, QUNIFORM_USER_CLIP_PLANE,
- plane * 4 + i);
- dist = qir_FADD(c, dist, qir_FMUL(c, pos_chan, ucp));
- }
-
- c->outputs[output_index] = dist;
- }
-}
-
static void
emit_vert_end(struct vc4_compile *c,
- struct vc4_varying_semantic *fs_inputs,
+ struct vc4_varying_slot *fs_inputs,
uint32_t num_fs_inputs)
{
struct qreg rcp_w = qir_RCP(c, c->outputs[c->output_position_index + 3]);
emit_stub_vpm_read(c);
- emit_ucp_clipdistance(c);
emit_scaled_viewport_write(c, rcp_w);
emit_zs_write(c, rcp_w);
emit_point_size_write(c);
for (int i = 0; i < num_fs_inputs; i++) {
- struct vc4_varying_semantic *input = &fs_inputs[i];
+ struct vc4_varying_slot *input = &fs_inputs[i];
int j;
for (j = 0; j < c->num_outputs; j++) {
- struct vc4_varying_semantic *output =
- &c->output_semantics[j];
+ struct vc4_varying_slot *output =
+ &c->output_slots[j];
- if (input->semantic == output->semantic &&
- input->index == output->index &&
+ if (input->slot == output->slot &&
input->swizzle == output->swizzle) {
qir_VPM_WRITE(c, c->outputs[j]);
break;
do {
progress = false;
- nir_lower_vars_to_ssa(s);
- nir_lower_alu_to_scalar(s);
+ NIR_PASS_V(s, nir_lower_vars_to_ssa);
+ NIR_PASS_V(s, nir_lower_alu_to_scalar);
- 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;
+ NIR_PASS(progress, s, nir_copy_prop);
+ NIR_PASS(progress, s, nir_opt_dce);
+ NIR_PASS(progress, s, nir_opt_cse);
+ NIR_PASS(progress, s, nir_opt_peephole_select);
+ NIR_PASS(progress, s, nir_opt_algebraic);
+ NIR_PASS(progress, s, nir_opt_constant_folding);
+ NIR_PASS(progress, s, nir_opt_undef);
} while (progress);
}
ntq_setup_inputs(struct vc4_compile *c)
{
unsigned num_entries = 0;
- foreach_list_typed(nir_variable, var, node, &c->s->inputs)
+ nir_foreach_variable(var, &c->s->inputs)
num_entries++;
nir_variable *vars[num_entries];
unsigned i = 0;
- foreach_list_typed(nir_variable, var, node, &c->s->inputs)
+ nir_foreach_variable(var, &c->s->inputs)
vars[i++] = var;
/* Sort the variables so that we emit the input setup in
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) {
+ if (var->data.location == VARYING_SLOT_POS) {
emit_fragcoord_input(c, loc);
- } else if (semantic_name == TGSI_SEMANTIC_FACE) {
- emit_face_input(c, loc);
- } else if (semantic_name == TGSI_SEMANTIC_GENERIC &&
+ } else if (var->data.location == VARYING_SLOT_FACE) {
+ c->inputs[loc * 4 + 0] = qir_FRAG_REV_FLAG(c);
+ } else if (var->data.location >= VARYING_SLOT_VAR0 &&
(c->fs_key->point_sprite_mask &
- (1 << semantic_index))) {
- emit_point_coord_input(c, loc);
+ (1 << (var->data.location -
+ VARYING_SLOT_VAR0)))) {
+ 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);
+ emit_fragment_input(c, loc, var->data.location);
}
} 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) {
+ nir_foreach_variable(var, &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;
- for (int i = 0; i < 4; i++) {
- add_output(c,
- loc + i,
- semantic_name,
- semantic_index,
- i);
- }
+ for (int i = 0; i < 4; i++)
+ add_output(c, loc + i, var->data.location, 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;
+ if (c->stage == QSTAGE_FRAG) {
+ switch (var->data.location) {
+ case FRAG_RESULT_COLOR:
+ case FRAG_RESULT_DATA0:
+ c->output_color_index = loc;
+ break;
+ case FRAG_RESULT_DEPTH:
+ c->output_position_index = loc;
+ break;
+ case FRAG_RESULT_SAMPLE_MASK:
+ c->output_sample_mask_index = loc;
+ break;
+ }
+ } else {
+ switch (var->data.location) {
+ case VARYING_SLOT_POS:
+ c->output_position_index = loc;
+ break;
+ case VARYING_SLOT_PSIZ:
+ 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) {
+ nir_foreach_variable(var, &c->s->uniforms) {
unsigned array_len = MAX2(glsl_get_length(var->type), 1);
unsigned array_elem_size = 4 * sizeof(float);
static void
ntq_emit_load_const(struct vc4_compile *c, nir_load_const_instr *instr)
{
- struct qreg *qregs = ralloc_array(c->def_ht, struct qreg,
- instr->def.num_components);
+ 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]);
+ qregs[i] = qir_uniform_ui(c, instr->value.u32[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];
+ nir_const_value *const_offset;
+ unsigned offset;
struct qreg *dest = NULL;
if (info->has_dest) {
- dest = ntq_get_dest(c, instr->dest);
+ dest = ntq_get_dest(c, &instr->dest);
}
switch (instr->intrinsic) {
case nir_intrinsic_load_uniform:
- for (int i = 0; i < instr->num_components; i++) {
- dest[i] = qir_uniform(c, QUNIFORM_UNIFORM,
- instr->const_index[0] * 4 + i);
+ assert(instr->num_components == 1);
+ const_offset = nir_src_as_const_value(instr->src[0]);
+ if (const_offset) {
+ offset = instr->const_index[0] + const_offset->u32[0];
+ assert(offset % 4 == 0);
+ /* We need dwords */
+ offset = offset / 4;
+ if (offset < VC4_NIR_STATE_UNIFORM_OFFSET) {
+ *dest = qir_uniform(c, QUNIFORM_UNIFORM,
+ offset);
+ } else {
+ *dest = qir_uniform(c, offset -
+ VC4_NIR_STATE_UNIFORM_OFFSET,
+ 0);
+ }
+ } else {
+ *dest = indirect_uniform_load(c, instr);
}
break;
- case nir_intrinsic_load_uniform_indirect:
+ case nir_intrinsic_load_user_clip_plane:
for (int i = 0; i < instr->num_components; i++) {
- dest[i] = indirect_uniform_load(c,
- ntq_get_src(c, instr->src[0], 0),
- (instr->const_index[0] *
- 4 + i) * sizeof(float));
+ dest[i] = qir_uniform(c, QUNIFORM_USER_CLIP_PLANE,
+ instr->const_index[0] * 4 + i);
}
+ break;
+ case nir_intrinsic_load_sample_mask_in:
+ *dest = qir_uniform(c, QUNIFORM_SAMPLE_MASK, 0);
break;
case nir_intrinsic_load_input:
- for (int i = 0; i < instr->num_components; i++)
- dest[i] = c->inputs[instr->const_index[0] * 4 + i];
-
+ assert(instr->num_components == 1);
+ const_offset = nir_src_as_const_value(instr->src[0]);
+ assert(const_offset && "vc4 doesn't support indirect inputs");
+ if (instr->const_index[0] >= VC4_NIR_TLB_COLOR_READ_INPUT) {
+ assert(const_offset->u32[0] == 0);
+ /* Reads of the per-sample color need to be done in
+ * order.
+ */
+ int sample_index = (instr->const_index[0] -
+ VC4_NIR_TLB_COLOR_READ_INPUT);
+ for (int i = 0; i <= sample_index; i++) {
+ if (c->color_reads[i].file == QFILE_NULL) {
+ c->color_reads[i] =
+ qir_TLB_COLOR_READ(c);
+ }
+ }
+ *dest = c->color_reads[sample_index];
+ } else {
+ offset = instr->const_index[0] + const_offset->u32[0];
+ *dest = c->inputs[offset];
+ }
break;
case nir_intrinsic_store_output:
- for (int i = 0; i < instr->num_components; i++) {
- c->outputs[instr->const_index[0] * 4 + i] =
- qir_MOV(c, ntq_get_src(c, instr->src[0], i));
+ const_offset = nir_src_as_const_value(instr->src[1]);
+ assert(const_offset && "vc4 doesn't support indirect outputs");
+ offset = instr->const_index[0] + const_offset->u32[0];
+
+ /* MSAA color outputs are the only case where we have an
+ * output that's not lowered to being a store of a single 32
+ * bit value.
+ */
+ if (c->stage == QSTAGE_FRAG && instr->num_components == 4) {
+ assert(offset == c->output_color_index);
+ for (int i = 0; i < 4; i++) {
+ c->sample_colors[i] =
+ qir_MOV(c, ntq_get_src(c, instr->src[0],
+ i));
+ }
+ } else {
+ assert(instr->num_components == 1);
+ c->outputs[offset] =
+ qir_MOV(c, ntq_get_src(c, instr->src[0], 0));
+ c->num_outputs = MAX2(c->num_outputs, offset + 1);
}
- c->num_outputs = MAX2(c->num_outputs,
- instr->const_index[0] * 4 +
- instr->num_components + 1);
break;
case nir_intrinsic_discard:
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;
}
}
+static void ntq_emit_cf_list(struct vc4_compile *c, struct exec_list *list);
+
+static void
+ntq_emit_loop(struct vc4_compile *c, nir_loop *nloop)
+{
+ fprintf(stderr, "LOOPS not fully handled. Rendering errors likely.\n");
+ ntq_emit_cf_list(c, &nloop->body);
+}
+
+static void
+ntq_emit_function(struct vc4_compile *c, nir_function_impl *func)
+{
+ fprintf(stderr, "FUNCTIONS not handled.\n");
+ abort();
+}
+
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;
ntq_emit_if(c, nir_cf_node_as_if(node));
break;
+ case nir_cf_node_loop:
+ ntq_emit_loop(c, nir_cf_node_as_loop(node));
+ break;
+
+ case nir_cf_node_function:
+ ntq_emit_function(c, nir_cf_node_as_function(node));
+ break;
+
default:
- assert(0);
+ fprintf(stderr, "Unknown NIR node type\n");
+ abort();
}
}
}
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);
+ nir_foreach_function(c->s, function) {
+ assert(strcmp(function->name, "main") == 0);
+ assert(function->impl);
+ ntq_emit_impl(c, function->impl);
}
}
static const nir_shader_compiler_options nir_options = {
+ .lower_extract_byte = true,
+ .lower_extract_word = true,
.lower_ffma = true,
.lower_flrp = true,
.lower_fpow = true,
count_nir_instrs(nir_shader *nir)
{
int count = 0;
- nir_foreach_overload(nir, overload) {
- if (!overload->impl)
+ nir_foreach_function(nir, function) {
+ if (!function->impl)
continue;
- nir_foreach_block(overload->impl, count_nir_instrs_in_block, &count);
+ nir_foreach_block(function->impl, count_nir_instrs_in_block, &count);
}
return count;
}
case QSTAGE_FRAG:
c->fs_key = (struct vc4_fs_key *)key;
if (c->fs_key->is_points) {
- c->point_x = emit_fragment_varying(c, ~0, ~0, 0);
- c->point_y = emit_fragment_varying(c, ~0, ~0, 0);
+ c->point_x = emit_fragment_varying(c, ~0, 0);
+ c->point_y = emit_fragment_varying(c, ~0, 0);
} else if (c->fs_key->is_lines) {
- c->line_x = emit_fragment_varying(c, ~0, ~0, 0);
+ c->line_x = emit_fragment_varying(c, ~0, 0);
}
break;
case QSTAGE_VERT:
}
const struct tgsi_token *tokens = key->shader_state->base.tokens;
- if (c->fs_key && c->fs_key->light_twoside) {
- if (!key->shader_state->twoside_tokens) {
- const struct tgsi_lowering_config lowering_config = {
- .color_two_side = true,
- };
- struct tgsi_shader_info info;
- key->shader_state->twoside_tokens =
- tgsi_transform_lowering(&lowering_config,
- key->shader_state->base.tokens,
- &info);
-
- /* If no transformation occurred, then NULL is
- * returned and we just use our original tokens.
- */
- if (!key->shader_state->twoside_tokens) {
- key->shader_state->twoside_tokens =
- key->shader_state->base.tokens;
- }
- }
- tokens = key->shader_state->twoside_tokens;
- }
if (vc4_debug & VC4_DEBUG_TGSI) {
fprintf(stderr, "%s prog %d/%d TGSI:\n",
}
c->s = tgsi_to_nir(tokens, &nir_options);
- nir_opt_global_to_local(c->s);
- nir_convert_to_ssa(c->s);
- nir_lower_idiv(c->s);
+ NIR_PASS_V(c->s, nir_opt_global_to_local);
+ NIR_PASS_V(c->s, nir_convert_to_ssa);
- vc4_optimize_nir(c->s);
+ if (stage == QSTAGE_FRAG)
+ NIR_PASS_V(c->s, vc4_nir_lower_blend, c);
+
+ struct nir_lower_tex_options tex_options = {
+ /* We would need to implement txs, but we don't want the
+ * int/float conversions
+ */
+ .lower_rect = false,
+
+ /* We want to use this, but we don't want to newton-raphson
+ * its rcp.
+ */
+ .lower_txp = false,
+
+ /* Apply swizzles to all samplers. */
+ .swizzle_result = ~0,
+ };
+
+ /* Lower the format swizzle and ARB_texture_swizzle-style swizzle.
+ * The format swizzling applies before sRGB decode, and
+ * ARB_texture_swizzle is the last thing before returning the sample.
+ */
+ for (int i = 0; i < ARRAY_SIZE(key->tex); i++) {
+ enum pipe_format format = c->key->tex[i].format;
+
+ if (!format)
+ continue;
+
+ const uint8_t *format_swizzle = vc4_get_format_swizzle(format);
+
+ for (int j = 0; j < 4; j++) {
+ uint8_t arb_swiz = c->key->tex[i].swizzle[j];
+
+ if (arb_swiz <= 3) {
+ tex_options.swizzles[i][j] =
+ format_swizzle[arb_swiz];
+ } else {
+ tex_options.swizzles[i][j] = arb_swiz;
+ }
+
+ /* If ARB_texture_swizzle is reading from the R, G, or
+ * B channels of an sRGB texture, then we need to
+ * apply sRGB decode to this channel at sample time.
+ */
+ if (arb_swiz < 3 && util_format_is_srgb(format)) {
+ c->tex_srgb_decode[i] |= (1 << j);
+ }
+
+ }
+ }
+
+ NIR_PASS_V(c->s, nir_lower_tex, &tex_options);
+
+ if (c->fs_key && c->fs_key->light_twoside)
+ NIR_PASS_V(c->s, nir_lower_two_sided_color);
+
+ if (stage == QSTAGE_FRAG)
+ NIR_PASS_V(c->s, nir_lower_clip_fs, c->key->ucp_enables);
+ else
+ NIR_PASS_V(c->s, nir_lower_clip_vs, c->key->ucp_enables);
- nir_remove_dead_variables(c->s);
+ NIR_PASS_V(c->s, vc4_nir_lower_io, c);
+ NIR_PASS_V(c->s, vc4_nir_lower_txf_ms, c);
+ NIR_PASS_V(c->s, nir_lower_idiv);
+ NIR_PASS_V(c->s, nir_lower_load_const_to_scalar);
- nir_convert_from_ssa(c->s);
+ vc4_optimize_nir(c->s);
+
+ NIR_PASS_V(c->s, nir_remove_dead_variables);
+ NIR_PASS_V(c->s, nir_convert_from_ssa, true);
if (vc4_debug & VC4_DEBUG_SHADERDB) {
fprintf(stderr, "SHADER-DB: %s prog %d/%d: %d NIR instructions\n",
break;
case QSTAGE_VERT:
emit_vert_end(c,
- vc4->prog.fs->input_semantics,
+ vc4->prog.fs->input_slots,
vc4->prog.fs->num_inputs);
break;
case QSTAGE_COORD:
qir_optimize(c);
qir_lower_uniforms(c);
+ qir_schedule_instructions(c);
+
if (vc4_debug & VC4_DEBUG_QIR) {
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);
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 *
shader->program_id = vc4->next_compiled_program_id++;
if (stage == QSTAGE_FRAG) {
- bool input_live[c->num_input_semantics];
- struct simple_node *node;
+ bool input_live[c->num_input_slots];
memset(input_live, 0, sizeof(input_live));
- foreach(node, &c->instructions) {
- struct qinst *inst = (struct qinst *)node;
+ 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);
+ shader->input_slots = ralloc_array(shader,
+ struct vc4_varying_slot,
+ c->num_input_slots);
- for (int i = 0; i < c->num_input_semantics; i++) {
- struct vc4_varying_semantic *sem = &c->input_semantics[i];
+ for (int i = 0; i < c->num_input_slots; i++) {
+ struct vc4_varying_slot *slot = &c->input_slots[i];
if (!input_live[i])
continue;
/* Skip non-VS-output inputs. */
- if (sem->semantic == (uint8_t)~0)
+ if (slot->slot == (uint8_t)~0)
continue;
- if (sem->semantic == TGSI_SEMANTIC_COLOR ||
- sem->semantic == TGSI_SEMANTIC_BCOLOR) {
+ if (slot->slot == VARYING_SLOT_COL0 ||
+ slot->slot == VARYING_SLOT_COL1 ||
+ slot->slot == VARYING_SLOT_BFC0 ||
+ slot->slot == VARYING_SLOT_BFC1) {
shader->color_inputs |= (1 << shader->num_inputs);
}
- shader->input_semantics[shader->num_inputs] = *sem;
+ shader->input_slots[shader->num_inputs] = *slot;
shader->num_inputs++;
}
} else {
}
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.
}
}
if (shader->ubo_size) {
- 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);
+ 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 pipe_sampler_state *sampler_state =
texstate->samplers[i];
- if (sampler) {
- key->tex[i].format = sampler->format;
- key->tex[i].swizzle[0] = sampler->swizzle_r;
- key->tex[i].swizzle[1] = sampler->swizzle_g;
- key->tex[i].swizzle[2] = sampler->swizzle_b;
- key->tex[i].swizzle[3] = sampler->swizzle_a;
+ if (!sampler)
+ continue;
+
+ key->tex[i].format = sampler->format;
+ key->tex[i].swizzle[0] = sampler->swizzle_r;
+ key->tex[i].swizzle[1] = sampler->swizzle_g;
+ key->tex[i].swizzle[2] = sampler->swizzle_b;
+ key->tex[i].swizzle[3] = sampler->swizzle_a;
+
+ if (sampler->texture->nr_samples > 1) {
+ key->tex[i].msaa_width = sampler->texture->width0;
+ key->tex[i].msaa_height = sampler->texture->height0;
+ } else if (sampler){
key->tex[i].compare_mode = sampler_state->compare_mode;
key->tex[i].compare_func = sampler_state->compare_func;
key->tex[i].wrap_s = sampler_state->wrap_s;
} else {
key->logicop_func = PIPE_LOGICOP_COPY;
}
+ key->msaa = vc4->rasterizer->base.multisample;
+ key->sample_coverage = (vc4->rasterizer->base.multisample &&
+ vc4->sample_mask != (1 << VC4_MAX_SAMPLES) - 1);
+ key->sample_alpha_to_coverage = vc4->blend->alpha_to_coverage;
+ key->sample_alpha_to_one = vc4->blend->alpha_to_one;
if (vc4->framebuffer.cbufs[0])
key->color_format = vc4->framebuffer.cbufs[0]->format;
(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
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((void *)so->base.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 -
- texture->u.tex.first_level, VC4_TEX_P0_MIPLVLS) |
- VC4_SET_FIELD(texture->target == PIPE_TEXTURE_CUBE,
- VC4_TEX_P0_CMMODE) |
- VC4_SET_FIELD(rsc->vc4_format & 15, 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_aligned_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_aligned_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_aligned_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);
-}
-
-static struct vc4_bo *
-vc4_upload_ubo(struct vc4_context *vc4, struct vc4_compiled_shader *shader,
- const uint32_t *gallium_uniforms)
-{
- if (!shader->ubo_size)
- return NULL;
-
- struct vc4_bo *ubo = vc4_bo_alloc(vc4->screen, shader->ubo_size, "ubo");
- uint32_t *data = vc4_bo_map(ubo);
- for (uint32_t i = 0; i < shader->num_ubo_ranges; i++) {
- memcpy(data + shader->ubo_ranges[i].dst_offset,
- gallium_uniforms + shader->ubo_ranges[i].src_offset,
- shader->ubo_ranges[i].size);
- }
-
- return ubo;
-}
-
-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;
- struct vc4_bo *ubo = vc4_upload_ubo(vc4, shader, gallium_uniforms);
-
- cl_ensure_space(&vc4->uniforms, (uinfo->count +
- uinfo->num_texture_samples) * 4);
-
- 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_aligned_u32(&vc4->uniforms, uinfo->data[i]);
- break;
- case QUNIFORM_UNIFORM:
- cl_aligned_u32(&vc4->uniforms,
- gallium_uniforms[uinfo->data[i]]);
- break;
- case QUNIFORM_VIEWPORT_X_SCALE:
- cl_aligned_f(&vc4->uniforms, vc4->viewport.scale[0] * 16.0f);
- break;
- case QUNIFORM_VIEWPORT_Y_SCALE:
- cl_aligned_f(&vc4->uniforms, vc4->viewport.scale[1] * 16.0f);
- break;
-
- case QUNIFORM_VIEWPORT_Z_OFFSET:
- cl_aligned_f(&vc4->uniforms, vc4->viewport.translate[2]);
- break;
- case QUNIFORM_VIEWPORT_Z_SCALE:
- cl_aligned_f(&vc4->uniforms, vc4->viewport.scale[2]);
- break;
-
- case QUNIFORM_USER_CLIP_PLANE:
- cl_aligned_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_UBO_ADDR:
- cl_aligned_reloc(vc4, &vc4->uniforms, ubo, 0);
- 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_aligned_u32(&vc4->uniforms,
- get_texrect_scale(texstate,
- uinfo->contents[i],
- uinfo->data[i]));
- break;
-
- case QUNIFORM_BLEND_CONST_COLOR:
- cl_aligned_f(&vc4->uniforms,
- CLAMP(vc4->blend_color.color[uinfo->data[i]], 0, 1));
- break;
-
- case QUNIFORM_STENCIL:
- cl_aligned_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_aligned_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)
{