#include "ir3_shader.h"
#include "ir3_compiler.h"
-
+#include "ir3_nir.h"
static void
delete_variant(struct ir3_shader_variant *v)
static void
assemble_variant(struct ir3_shader_variant *v)
{
- struct fd_context *ctx = fd_context(v->shader->pctx);
- uint32_t gpu_id = v->shader->compiler->gpu_id;
+ struct ir3_compiler *compiler = v->shader->compiler;
+ uint32_t gpu_id = compiler->gpu_id;
uint32_t sz, *bin;
bin = ir3_shader_assemble(v, gpu_id);
sz = v->info.sizedwords * 4;
- v->bo = fd_bo_new(ctx->dev, sz,
+ v->bo = fd_bo_new(compiler->dev, sz,
DRM_FREEDRENO_GEM_CACHE_WCOMBINE |
DRM_FREEDRENO_GEM_TYPE_KMEM);
create_variant(struct ir3_shader *shader, struct ir3_shader_key key)
{
struct ir3_shader_variant *v = CALLOC_STRUCT(ir3_shader_variant);
- const struct tgsi_token *tokens = shader->tokens;
int ret;
if (!v)
v->key = key;
v->type = shader->type;
- if (fd_mesa_debug & FD_DBG_DISASM) {
- DBG("dump tgsi: type=%d, k={bp=%u,cts=%u,hp=%u}", shader->type,
- key.binning_pass, key.color_two_side, key.half_precision);
- tgsi_dump(tokens, 0);
- }
-
- ret = ir3_compile_shader_nir(shader->compiler, v, tokens, key);
+ ret = ir3_compile_shader_nir(shader->compiler, v);
if (ret) {
debug_error("compile failed!");
goto fail;
key.vsaturate_s = 0;
key.vsaturate_t = 0;
key.vsaturate_r = 0;
+ key.vastc_srgb = 0;
}
break;
case SHADER_VERTEX:
key.fsaturate_s = 0;
key.fsaturate_t = 0;
key.fsaturate_r = 0;
+ key.fastc_srgb = 0;
}
break;
}
v = v->next;
delete_variant(t);
}
- free((void *)shader->tokens);
+ ralloc_free(shader->nir);
free(shader);
}
struct ir3_shader *
-ir3_shader_create(struct pipe_context *pctx, const struct tgsi_token *tokens,
+ir3_shader_create(struct ir3_compiler *compiler,
+ const struct pipe_shader_state *cso,
enum shader_t type)
{
struct ir3_shader *shader = CALLOC_STRUCT(ir3_shader);
- shader->compiler = fd_context(pctx)->screen->compiler;
+ shader->compiler = compiler;
shader->id = ++shader->compiler->shader_count;
- shader->pctx = pctx;
shader->type = type;
- shader->tokens = tgsi_dup_tokens(tokens);
+ if (fd_mesa_debug & FD_DBG_DISASM) {
+ DBG("dump tgsi: type=%d", shader->type);
+ tgsi_dump(cso->tokens, 0);
+ }
+ nir_shader *nir = ir3_tgsi_to_nir(cso->tokens);
+ /* do first pass optimization, ignoring the key: */
+ shader->nir = ir3_optimize_nir(shader, nir, NULL);
+ if (fd_mesa_debug & FD_DBG_DISASM) {
+ DBG("dump nir%d: type=%d", shader->id, shader->type);
+ nir_print_shader(shader->nir, stdout);
+ }
+ shader->stream_output = cso->stream_output;
if (fd_mesa_debug & FD_DBG_SHADERDB) {
/* if shader-db run, create a standard variant immediately
* (as otherwise nothing will trigger the shader to be
* actually compiled)
*/
- static struct ir3_shader_key key = {};
+ static struct ir3_shader_key key = {0};
ir3_shader_variant(shader, key);
}
return shader;
debug_printf("; %s: r%d.%c\n", name, r >> 2, "xyzw"[r & 0x3]);
}
-static void dump_semantic(struct ir3_shader_variant *so,
- unsigned sem, const char *name)
+static void dump_output(struct ir3_shader_variant *so,
+ unsigned slot, const char *name)
{
uint32_t regid;
- regid = ir3_find_output_regid(so, ir3_semantic_name(sem, 0));
+ regid = ir3_find_output_regid(so, slot);
dump_reg(name, regid);
}
disasm_a3xx(bin, so->info.sizedwords, 0, so->type);
- debug_printf("; %s: outputs:", type);
- for (i = 0; i < so->outputs_count; i++) {
- uint8_t regid = so->outputs[i].regid;
- ir3_semantic sem = so->outputs[i].semantic;
- debug_printf(" r%d.%c (%u:%u)",
- (regid >> 2), "xyzw"[regid & 0x3],
- sem2name(sem), sem2idx(sem));
- }
- debug_printf("\n");
- debug_printf("; %s: inputs:", type);
- for (i = 0; i < so->inputs_count; i++) {
- uint8_t regid = so->inputs[i].regid;
- ir3_semantic sem = so->inputs[i].semantic;
- debug_printf(" r%d.%c (%u:%u,cm=%x,il=%u,b=%u)",
- (regid >> 2), "xyzw"[regid & 0x3],
- sem2name(sem), sem2idx(sem),
- so->inputs[i].compmask,
- so->inputs[i].inloc,
- so->inputs[i].bary);
+ switch (so->type) {
+ case SHADER_VERTEX:
+ debug_printf("; %s: outputs:", type);
+ for (i = 0; i < so->outputs_count; i++) {
+ uint8_t regid = so->outputs[i].regid;
+ debug_printf(" r%d.%c (%s)",
+ (regid >> 2), "xyzw"[regid & 0x3],
+ gl_varying_slot_name(so->outputs[i].slot));
+ }
+ debug_printf("\n");
+ debug_printf("; %s: inputs:", type);
+ for (i = 0; i < so->inputs_count; i++) {
+ uint8_t regid = so->inputs[i].regid;
+ debug_printf(" r%d.%c (cm=%x,il=%u,b=%u)",
+ (regid >> 2), "xyzw"[regid & 0x3],
+ so->inputs[i].compmask,
+ so->inputs[i].inloc,
+ so->inputs[i].bary);
+ }
+ debug_printf("\n");
+ break;
+ case SHADER_FRAGMENT:
+ debug_printf("; %s: outputs:", type);
+ for (i = 0; i < so->outputs_count; i++) {
+ uint8_t regid = so->outputs[i].regid;
+ debug_printf(" r%d.%c (%s)",
+ (regid >> 2), "xyzw"[regid & 0x3],
+ gl_frag_result_name(so->outputs[i].slot));
+ }
+ debug_printf("\n");
+ debug_printf("; %s: inputs:", type);
+ for (i = 0; i < so->inputs_count; i++) {
+ uint8_t regid = so->inputs[i].regid;
+ debug_printf(" r%d.%c (%s,cm=%x,il=%u,b=%u)",
+ (regid >> 2), "xyzw"[regid & 0x3],
+ gl_varying_slot_name(so->inputs[i].slot),
+ so->inputs[i].compmask,
+ so->inputs[i].inloc,
+ so->inputs[i].bary);
+ }
+ debug_printf("\n");
+ break;
+ case SHADER_COMPUTE:
+ break;
}
- debug_printf("\n");
/* print generic shader info: */
debug_printf("; %s prog %d/%d: %u instructions, %d half, %d full\n",
/* print shader type specific info: */
switch (so->type) {
case SHADER_VERTEX:
- dump_semantic(so, TGSI_SEMANTIC_POSITION, "pos");
- dump_semantic(so, TGSI_SEMANTIC_PSIZE, "psize");
+ dump_output(so, VARYING_SLOT_POS, "pos");
+ dump_output(so, VARYING_SLOT_PSIZ, "psize");
break;
case SHADER_FRAGMENT:
dump_reg("pos (bary)", so->pos_regid);
- dump_semantic(so, TGSI_SEMANTIC_POSITION, "posz");
- dump_semantic(so, TGSI_SEMANTIC_COLOR, "color");
+ dump_output(so, FRAG_RESULT_DEPTH, "posz");
+ if (so->color0_mrt) {
+ dump_output(so, FRAG_RESULT_COLOR, "color");
+ } else {
+ dump_output(so, FRAG_RESULT_DATA0, "data0");
+ dump_output(so, FRAG_RESULT_DATA1, "data1");
+ dump_output(so, FRAG_RESULT_DATA2, "data2");
+ dump_output(so, FRAG_RESULT_DATA3, "data3");
+ dump_output(so, FRAG_RESULT_DATA4, "data4");
+ dump_output(so, FRAG_RESULT_DATA5, "data5");
+ dump_output(so, FRAG_RESULT_DATA6, "data6");
+ dump_output(so, FRAG_RESULT_DATA7, "data7");
+ }
/* these two are hard-coded since we don't know how to
* program them to anything but all 0's...
*/
debug_printf("\n");
}
+
+/* This has to reach into the fd_context a bit more than the rest of
+ * ir3, but it needs to be aligned with the compiler, so both agree
+ * on which const regs hold what. And the logic is identical between
+ * a3xx/a4xx, the only difference is small details in the actual
+ * CP_LOAD_STATE packets (which is handled inside the generation
+ * specific ctx->emit_const(_bo)() fxns)
+ */
+
+#include "freedreno_resource.h"
+
+static void
+emit_user_consts(struct fd_context *ctx, const struct ir3_shader_variant *v,
+ struct fd_ringbuffer *ring, struct fd_constbuf_stateobj *constbuf)
+{
+ const unsigned index = 0; /* user consts are index 0 */
+ /* TODO save/restore dirty_mask for binning pass instead: */
+ uint32_t dirty_mask = constbuf->enabled_mask;
+
+ if (dirty_mask & (1 << index)) {
+ struct pipe_constant_buffer *cb = &constbuf->cb[index];
+ unsigned size = align(cb->buffer_size, 4) / 4; /* size in dwords */
+
+ /* in particular, with binning shader we may end up with
+ * unused consts, ie. we could end up w/ constlen that is
+ * smaller than first_driver_param. In that case truncate
+ * the user consts early to avoid HLSQ lockup caused by
+ * writing too many consts
+ */
+ uint32_t max_const = MIN2(v->first_driver_param, v->constlen);
+
+ // I expect that size should be a multiple of vec4's:
+ assert(size == align(size, 4));
+
+ /* and even if the start of the const buffer is before
+ * first_immediate, the end may not be:
+ */
+ size = MIN2(size, 4 * max_const);
+
+ if (size > 0) {
+ fd_wfi(ctx, ring);
+ ctx->emit_const(ring, v->type, 0,
+ cb->buffer_offset, size,
+ cb->user_buffer, cb->buffer);
+ constbuf->dirty_mask &= ~(1 << index);
+ }
+ }
+}
+
+static void
+emit_ubos(struct fd_context *ctx, const struct ir3_shader_variant *v,
+ struct fd_ringbuffer *ring, struct fd_constbuf_stateobj *constbuf)
+{
+ uint32_t offset = v->first_driver_param + IR3_UBOS_OFF;
+ if (v->constlen > offset) {
+ uint32_t params = MIN2(4, v->constlen - offset) * 4;
+ uint32_t offsets[params];
+ struct fd_bo *bos[params];
+
+ for (uint32_t i = 0; i < params; i++) {
+ const uint32_t index = i + 1; /* UBOs start at index 1 */
+ struct pipe_constant_buffer *cb = &constbuf->cb[index];
+ assert(!cb->user_buffer);
+
+ if ((constbuf->enabled_mask & (1 << index)) && cb->buffer) {
+ offsets[i] = cb->buffer_offset;
+ bos[i] = fd_resource(cb->buffer)->bo;
+ } else {
+ offsets[i] = 0;
+ bos[i] = NULL;
+ }
+ }
+
+ fd_wfi(ctx, ring);
+ ctx->emit_const_bo(ring, v->type, false, offset * 4, params, bos, offsets);
+ }
+}
+
+static void
+emit_immediates(struct fd_context *ctx, const struct ir3_shader_variant *v,
+ struct fd_ringbuffer *ring)
+{
+ int size = v->immediates_count;
+ uint32_t base = v->first_immediate;
+
+ /* truncate size to avoid writing constants that shader
+ * does not use:
+ */
+ size = MIN2(size + base, v->constlen) - base;
+
+ /* convert out of vec4: */
+ base *= 4;
+ size *= 4;
+
+ if (size > 0) {
+ fd_wfi(ctx, ring);
+ ctx->emit_const(ring, v->type, base,
+ 0, size, v->immediates[0].val, NULL);
+ }
+}
+
+/* emit stream-out buffers: */
+static void
+emit_tfbos(struct fd_context *ctx, const struct ir3_shader_variant *v,
+ struct fd_ringbuffer *ring)
+{
+ /* streamout addresses after driver-params: */
+ uint32_t offset = v->first_driver_param + IR3_TFBOS_OFF;
+ if (v->constlen > offset) {
+ struct fd_streamout_stateobj *so = &ctx->streamout;
+ struct pipe_stream_output_info *info = &v->shader->stream_output;
+ uint32_t params = 4;
+ uint32_t offsets[params];
+ struct fd_bo *bos[params];
+
+ for (uint32_t i = 0; i < params; i++) {
+ struct pipe_stream_output_target *target = so->targets[i];
+
+ if (target) {
+ offsets[i] = (so->offsets[i] * info->stride[i] * 4) +
+ target->buffer_offset;
+ bos[i] = fd_resource(target->buffer)->bo;
+ } else {
+ offsets[i] = 0;
+ bos[i] = NULL;
+ }
+ }
+
+ fd_wfi(ctx, ring);
+ ctx->emit_const_bo(ring, v->type, true, offset * 4, params, bos, offsets);
+ }
+}
+
+static uint32_t
+max_tf_vtx(struct fd_context *ctx, const struct ir3_shader_variant *v)
+{
+ struct fd_streamout_stateobj *so = &ctx->streamout;
+ struct pipe_stream_output_info *info = &v->shader->stream_output;
+ uint32_t maxvtxcnt = 0x7fffffff;
+
+ if (v->key.binning_pass)
+ return 0;
+ if (v->shader->stream_output.num_outputs == 0)
+ return 0;
+ if (so->num_targets == 0)
+ return 0;
+
+ /* offset to write to is:
+ *
+ * total_vtxcnt = vtxcnt + offsets[i]
+ * offset = total_vtxcnt * stride[i]
+ *
+ * offset = vtxcnt * stride[i] ; calculated in shader
+ * + offsets[i] * stride[i] ; calculated at emit_tfbos()
+ *
+ * assuming for each vtx, each target buffer will have data written
+ * up to 'offset + stride[i]', that leaves maxvtxcnt as:
+ *
+ * buffer_size = (maxvtxcnt * stride[i]) + stride[i]
+ * maxvtxcnt = (buffer_size - stride[i]) / stride[i]
+ *
+ * but shader is actually doing a less-than (rather than less-than-
+ * equal) check, so we can drop the -stride[i].
+ *
+ * TODO is assumption about `offset + stride[i]` legit?
+ */
+ for (unsigned i = 0; i < so->num_targets; i++) {
+ struct pipe_stream_output_target *target = so->targets[i];
+ unsigned stride = info->stride[i] * 4; /* convert dwords->bytes */
+ if (target) {
+ uint32_t max = target->buffer_size / stride;
+ maxvtxcnt = MIN2(maxvtxcnt, max);
+ }
+ }
+
+ return maxvtxcnt;
+}
+
+void
+ir3_emit_consts(const struct ir3_shader_variant *v, struct fd_ringbuffer *ring,
+ struct fd_context *ctx, const struct pipe_draw_info *info, uint32_t dirty)
+{
+ if (dirty & (FD_DIRTY_PROG | FD_DIRTY_CONSTBUF)) {
+ struct fd_constbuf_stateobj *constbuf;
+ bool shader_dirty;
+
+ if (v->type == SHADER_VERTEX) {
+ constbuf = &ctx->constbuf[PIPE_SHADER_VERTEX];
+ shader_dirty = !!(ctx->prog.dirty & FD_SHADER_DIRTY_VP);
+ } else if (v->type == SHADER_FRAGMENT) {
+ constbuf = &ctx->constbuf[PIPE_SHADER_FRAGMENT];
+ shader_dirty = !!(ctx->prog.dirty & FD_SHADER_DIRTY_FP);
+ } else {
+ unreachable("bad shader type");
+ return;
+ }
+
+ emit_user_consts(ctx, v, ring, constbuf);
+ emit_ubos(ctx, v, ring, constbuf);
+ if (shader_dirty)
+ emit_immediates(ctx, v, ring);
+ }
+
+ /* emit driver params every time: */
+ /* TODO skip emit if shader doesn't use driver params to avoid WFI.. */
+ if (info && (v->type == SHADER_VERTEX)) {
+ uint32_t offset = v->first_driver_param + IR3_DRIVER_PARAM_OFF;
+ if (v->constlen >= offset) {
+ uint32_t vertex_params[IR3_DP_COUNT] = {
+ [IR3_DP_VTXID_BASE] = info->indexed ?
+ info->index_bias : info->start,
+ [IR3_DP_VTXCNT_MAX] = max_tf_vtx(ctx, v),
+ };
+ /* if no user-clip-planes, we don't need to emit the
+ * entire thing:
+ */
+ uint32_t vertex_params_size = 4;
+
+ if (v->key.ucp_enables) {
+ struct pipe_clip_state *ucp = &ctx->ucp;
+ unsigned pos = IR3_DP_UCP0_X;
+ for (unsigned i = 0; pos <= IR3_DP_UCP7_W; i++) {
+ for (unsigned j = 0; j < 4; j++) {
+ vertex_params[pos] = fui(ucp->ucp[i][j]);
+ pos++;
+ }
+ }
+ vertex_params_size = ARRAY_SIZE(vertex_params);
+ }
+
+ fd_wfi(ctx, ring);
+ ctx->emit_const(ring, SHADER_VERTEX, offset * 4, 0,
+ vertex_params_size, vertex_params, NULL);
+
+ /* if needed, emit stream-out buffer addresses: */
+ if (vertex_params[IR3_DP_VTXCNT_MAX] > 0) {
+ emit_tfbos(ctx, v, ring);
+ }
+ }
+ }
+}