#include "lp_tex_sample.h"
#include "lp_flush.h"
#include "lp_state_fs.h"
+#include "lp_rast.h"
/** Fragment shader number (for debugging) */
}
-/**
- * Generate the fragment shader, depth/stencil test, and alpha tests.
- * \param i which quad in the tile, in range [0,3]
- * \param partial_mask if 1, do mask_input testing
- */
-static void
-generate_fs(struct gallivm_state *gallivm,
- struct lp_fragment_shader *shader,
- const struct lp_fragment_shader_variant_key *key,
- LLVMBuilderRef builder,
- struct lp_type type,
- LLVMValueRef context_ptr,
- unsigned i,
- struct lp_build_interp_soa_context *interp,
- struct lp_build_sampler_soa *sampler,
- LLVMValueRef *pmask,
- LLVMValueRef (*color)[4],
- LLVMValueRef depth_ptr,
- LLVMValueRef facing,
- unsigned partial_mask,
- LLVMValueRef mask_input,
- LLVMValueRef thread_data_ptr)
-{
- const struct util_format_description *zs_format_desc = NULL;
- const struct tgsi_token *tokens = shader->base.tokens;
- LLVMTypeRef vec_type;
- LLVMValueRef consts_ptr;
- LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
- LLVMValueRef z;
- LLVMValueRef zs_value = NULL;
- LLVMValueRef stencil_refs[2];
- struct lp_build_mask_context mask;
- boolean simple_shader = (shader->info.base.file_count[TGSI_FILE_SAMPLER] == 0 &&
- shader->info.base.num_inputs < 3 &&
- shader->info.base.num_instructions < 8);
- const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
- util_blend_state_is_dual(&key->blend, 0);
- unsigned attrib;
- unsigned chan;
- unsigned cbuf;
- unsigned depth_mode;
- struct lp_bld_tgsi_system_values system_values;
-
- memset(&system_values, 0, sizeof(system_values));
-
- if (key->depth.enabled ||
- key->stencil[0].enabled ||
- key->stencil[1].enabled) {
-
- zs_format_desc = util_format_description(key->zsbuf_format);
- assert(zs_format_desc);
-
- if (!shader->info.base.writes_z) {
- if (key->alpha.enabled || shader->info.base.uses_kill)
- /* With alpha test and kill, can do the depth test early
- * and hopefully eliminate some quads. But need to do a
- * special deferred depth write once the final mask value
- * is known.
- */
- depth_mode = EARLY_DEPTH_TEST | LATE_DEPTH_WRITE;
- else
- depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
- }
- else {
- depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
- }
-
- if (!(key->depth.enabled && key->depth.writemask) &&
- !(key->stencil[0].enabled && key->stencil[0].writemask))
- depth_mode &= ~(LATE_DEPTH_WRITE | EARLY_DEPTH_WRITE);
- }
- else {
- depth_mode = 0;
- }
-
- assert(i < 4);
-
- stencil_refs[0] = lp_jit_context_stencil_ref_front_value(gallivm, context_ptr);
- stencil_refs[1] = lp_jit_context_stencil_ref_back_value(gallivm, context_ptr);
-
- vec_type = lp_build_vec_type(gallivm, type);
-
- consts_ptr = lp_jit_context_constants(gallivm, context_ptr);
-
- memset(outputs, 0, sizeof outputs);
-
- /* Declare the color and z variables */
- for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
- for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
- color[cbuf][chan] = lp_build_alloca(gallivm, vec_type, "color");
- }
- }
- if (dual_source_blend) {
- assert(key->nr_cbufs <= 1);
- for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
- color[1][chan] = lp_build_alloca(gallivm, vec_type, "color1");
- }
- }
-
- /* do triangle edge testing */
- if (partial_mask) {
- *pmask = generate_quad_mask(gallivm, type,
- i*type.length/4, mask_input);
- }
- else {
- *pmask = lp_build_const_int_vec(gallivm, type, ~0);
- }
-
- /* 'mask' will control execution based on quad's pixel alive/killed state */
- lp_build_mask_begin(&mask, gallivm, type, *pmask);
-
- if (!(depth_mode & EARLY_DEPTH_TEST) && !simple_shader)
- lp_build_mask_check(&mask);
-
- lp_build_interp_soa_update_pos(interp, gallivm, i*type.length/4);
- z = interp->pos[2];
-
- if (depth_mode & EARLY_DEPTH_TEST) {
- lp_build_depth_stencil_test(gallivm,
- &key->depth,
- key->stencil,
- type,
- zs_format_desc,
- &mask,
- stencil_refs,
- z,
- depth_ptr, facing,
- &zs_value,
- !simple_shader);
-
- if (depth_mode & EARLY_DEPTH_WRITE) {
- lp_build_depth_write(gallivm, type, zs_format_desc, depth_ptr, zs_value);
- }
- }
-
- lp_build_interp_soa_update_inputs(interp, gallivm, i*type.length/4);
-
- /* Build the actual shader */
- lp_build_tgsi_soa(gallivm, tokens, type, &mask,
- consts_ptr, &system_values,
- interp->pos, interp->inputs,
- outputs, sampler, &shader->info.base);
-
- /* Alpha test */
- if (key->alpha.enabled) {
- int color0 = find_output_by_semantic(&shader->info.base,
- TGSI_SEMANTIC_COLOR,
- 0);
-
- if (color0 != -1 && outputs[color0][3]) {
- const struct util_format_description *cbuf_format_desc;
- LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
- LLVMValueRef alpha_ref_value;
-
- alpha_ref_value = lp_jit_context_alpha_ref_value(gallivm, context_ptr);
- alpha_ref_value = lp_build_broadcast(gallivm, vec_type, alpha_ref_value);
-
- cbuf_format_desc = util_format_description(key->cbuf_format[0]);
-
- lp_build_alpha_test(gallivm, key->alpha.func, type, cbuf_format_desc,
- &mask, alpha, alpha_ref_value,
- (depth_mode & LATE_DEPTH_TEST) != 0);
- }
- }
-
- /* Late Z test */
- if (depth_mode & LATE_DEPTH_TEST) {
- int pos0 = find_output_by_semantic(&shader->info.base,
- TGSI_SEMANTIC_POSITION,
- 0);
-
- if (pos0 != -1 && outputs[pos0][2]) {
- z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
- }
-
- lp_build_depth_stencil_test(gallivm,
- &key->depth,
- key->stencil,
- type,
- zs_format_desc,
- &mask,
- stencil_refs,
- z,
- depth_ptr, facing,
- &zs_value,
- !simple_shader);
- /* Late Z write */
- if (depth_mode & LATE_DEPTH_WRITE) {
- lp_build_depth_write(gallivm, type, zs_format_desc, depth_ptr, zs_value);
- }
- }
- else if ((depth_mode & EARLY_DEPTH_TEST) &&
- (depth_mode & LATE_DEPTH_WRITE))
- {
- /* Need to apply a reduced mask to the depth write. Reload the
- * depth value, update from zs_value with the new mask value and
- * write that out.
- */
- lp_build_deferred_depth_write(gallivm,
- type,
- zs_format_desc,
- &mask,
- depth_ptr,
- zs_value);
- }
-
-
- /* Color write */
- for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
- {
- unsigned cbuf = shader->info.base.output_semantic_index[attrib];
- if ((shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR) &&
- ((cbuf < key->nr_cbufs) || (cbuf == 1 && dual_source_blend)))
- {
- unsigned cbuf = shader->info.base.output_semantic_index[attrib];
- for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
- if(outputs[attrib][chan]) {
- /* XXX: just initialize outputs to point at colors[] and
- * skip this.
- */
- LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
- lp_build_name(out, "color%u.%u.%c", i, attrib, "rgba"[chan]);
- LLVMBuildStore(builder, out, color[cbuf][chan]);
- }
- }
- }
- }
-
- if (key->occlusion_count) {
- LLVMValueRef counter = lp_jit_thread_data_counter(gallivm, thread_data_ptr);
- lp_build_name(counter, "counter");
- lp_build_occlusion_count(gallivm, type,
- lp_build_mask_value(&mask), counter);
- }
-
- *pmask = lp_build_mask_end(&mask);
-}
-
-
/**
* Generate the fragment shader, depth/stencil test, and alpha tests.
*/
LLVMValueRef mask_store,
LLVMValueRef (*out_color)[4],
LLVMValueRef depth_ptr,
- unsigned depth_bits,
+ LLVMValueRef depth_stride,
LLVMValueRef facing,
LLVMValueRef thread_data_ptr)
{
LLVMValueRef mask_ptr, mask_val;
LLVMValueRef consts_ptr;
LLVMValueRef z;
- LLVMValueRef zs_value = NULL;
+ LLVMValueRef z_value, s_value;
+ LLVMValueRef z_fb, s_fb;
LLVMValueRef stencil_refs[2];
- LLVMValueRef depth_ptr_i;
- LLVMValueRef depth_offset;
LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
struct lp_build_for_loop_state loop_state;
struct lp_build_mask_context mask;
+ /*
+ * TODO: figure out if simple_shader optimization is really worthwile to
+ * keep. Disabled because it may hide some real bugs in the (depth/stencil)
+ * code since tests tend to take another codepath than real shaders.
+ */
boolean simple_shader = (shader->info.base.file_count[TGSI_FILE_SAMPLER] == 0 &&
shader->info.base.num_inputs < 3 &&
- shader->info.base.num_instructions < 8);
+ shader->info.base.num_instructions < 8) && 0;
const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
util_blend_state_is_dual(&key->blend, 0);
unsigned attrib;
memset(&system_values, 0, sizeof(system_values));
if (key->depth.enabled ||
- key->stencil[0].enabled ||
- key->stencil[1].enabled) {
+ key->stencil[0].enabled) {
zs_format_desc = util_format_description(key->zsbuf_format);
assert(zs_format_desc);
if (!shader->info.base.writes_z) {
- if (key->alpha.enabled || shader->info.base.uses_kill)
+ if (key->alpha.enabled || shader->info.base.uses_kill) {
/* With alpha test and kill, can do the depth test early
* and hopefully eliminate some quads. But need to do a
* special deferred depth write once the final mask value
- * is known.
+ * is known. This only works though if there's either no
+ * stencil test or the stencil value isn't written.
*/
- depth_mode = EARLY_DEPTH_TEST | LATE_DEPTH_WRITE;
+ if (key->stencil[0].enabled && (key->stencil[0].writemask ||
+ (key->stencil[1].enabled &&
+ key->stencil[1].writemask)))
+ depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
+ else
+ depth_mode = EARLY_DEPTH_TEST | LATE_DEPTH_WRITE;
+ }
else
depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
}
}
if (!(key->depth.enabled && key->depth.writemask) &&
- !(key->stencil[0].enabled && key->stencil[0].writemask))
+ !(key->stencil[0].enabled && (key->stencil[0].writemask ||
+ (key->stencil[1].enabled &&
+ key->stencil[1].writemask))))
depth_mode &= ~(LATE_DEPTH_WRITE | EARLY_DEPTH_WRITE);
}
else {
&loop_state.counter, 1, "mask_ptr");
mask_val = LLVMBuildLoad(builder, mask_ptr, "");
- depth_offset = LLVMBuildMul(builder, loop_state.counter,
- lp_build_const_int32(gallivm, depth_bits * type.length),
- "");
-
- depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &depth_offset, 1, "");
-
memset(outputs, 0, sizeof outputs);
for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
z = interp->pos[2];
if (depth_mode & EARLY_DEPTH_TEST) {
+ lp_build_depth_stencil_load_swizzled(gallivm, type,
+ zs_format_desc, key->resource_1d,
+ depth_ptr, depth_stride,
+ &z_fb, &s_fb, loop_state.counter);
lp_build_depth_stencil_test(gallivm,
&key->depth,
key->stencil,
zs_format_desc,
&mask,
stencil_refs,
- z,
- depth_ptr_i, facing,
- &zs_value,
+ z, z_fb, s_fb,
+ facing,
+ &z_value, &s_value,
!simple_shader);
if (depth_mode & EARLY_DEPTH_WRITE) {
- lp_build_depth_write(gallivm, type, zs_format_desc, depth_ptr_i, zs_value);
+ lp_build_depth_stencil_write_swizzled(gallivm, type,
+ zs_format_desc, key->resource_1d,
+ NULL, NULL, NULL, loop_state.counter,
+ depth_ptr, depth_stride,
+ z_value, s_value);
}
+ /*
+ * Note mask check if stencil is enabled must be after ds write not after
+ * stencil test otherwise new stencil values may not get written if all
+ * fragments got killed by depth/stencil test.
+ */
+ if (!simple_shader && key->stencil[0].enabled)
+ lp_build_mask_check(&mask);
}
lp_build_interp_soa_update_inputs_dyn(interp, gallivm, loop_state.counter);
/* Build the actual shader */
lp_build_tgsi_soa(gallivm, tokens, type, &mask,
consts_ptr, &system_values,
- interp->pos, interp->inputs,
- outputs, sampler, &shader->info.base);
+ interp->inputs,
+ outputs, sampler, &shader->info.base, NULL);
/* Alpha test */
if (key->alpha.enabled) {
z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
}
+ lp_build_depth_stencil_load_swizzled(gallivm, type,
+ zs_format_desc, key->resource_1d,
+ depth_ptr, depth_stride,
+ &z_fb, &s_fb, loop_state.counter);
+
lp_build_depth_stencil_test(gallivm,
&key->depth,
key->stencil,
zs_format_desc,
&mask,
stencil_refs,
- z,
- depth_ptr_i, facing,
- &zs_value,
+ z, z_fb, s_fb,
+ facing,
+ &z_value, &s_value,
!simple_shader);
/* Late Z write */
if (depth_mode & LATE_DEPTH_WRITE) {
- lp_build_depth_write(gallivm, type, zs_format_desc, depth_ptr_i, zs_value);
+ lp_build_depth_stencil_write_swizzled(gallivm, type,
+ zs_format_desc, key->resource_1d,
+ NULL, NULL, NULL, loop_state.counter,
+ depth_ptr, depth_stride,
+ z_value, s_value);
}
}
else if ((depth_mode & EARLY_DEPTH_TEST) &&
* depth value, update from zs_value with the new mask value and
* write that out.
*/
- lp_build_deferred_depth_write(gallivm,
- type,
- zs_format_desc,
- &mask,
- depth_ptr_i,
- zs_value);
+ lp_build_depth_stencil_write_swizzled(gallivm, type,
+ zs_format_desc, key->resource_1d,
+ &mask, z_fb, s_fb, loop_state.counter,
+ depth_ptr, depth_stride,
+ z_value, s_value);
}
*
* @param type fragment shader type (4x or 8x float)
* @param num_fs number of fs_src
+ * @param is_1d whether we're outputting to a 1d resource
* @param dst_channels number of output channels
* @param fs_src output from fragment shader
* @param dst pointer to store result
bool twiddle;
bool split;
- unsigned pixels = num_fs == 4 ? 1 : 2;
+ unsigned pixels = type.length / 4;
unsigned reorder_group;
unsigned src_channels;
unsigned src_count;
src_channels = dst_channels < 3 ? dst_channels : 4;
src_count = num_fs * src_channels;
- assert(pixels == 2 || num_fs == 4);
+ assert(pixels == 2 || pixels == 1);
assert(num_fs * src_channels <= Elements(src));
/*
}
+/**
+ * Checks if this format requires special handling due to required expansion
+ * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
+ * SoA conversion.
+ */
+static INLINE boolean
+format_expands_to_float_soa(const struct util_format_description *format_desc)
+{
+ if (format_desc->format == PIPE_FORMAT_R11G11B10_FLOAT ||
+ format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
+ return true;
+ }
+ return false;
+}
+
+
/**
* Retrieves the type representing the memory layout for a format
*
unsigned i;
unsigned chan;
+ if (format_expands_to_float_soa(format_desc)) {
+ /* just make this a 32bit uint */
+ type->floating = false;
+ type->fixed = false;
+ type->sign = false;
+ type->norm = false;
+ type->width = 32;
+ type->length = 1;
+ return;
+ }
+
for (i = 0; i < 4; i++)
if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
break;
unsigned i;
unsigned chan;
+ if (format_expands_to_float_soa(format_desc)) {
+ /* always use ordinary floats for blending */
+ type->floating = true;
+ type->fixed = false;
+ type->sign = true;
+ type->norm = false;
+ type->width = 32;
+ type->length = 4;
+ return;
+ }
+
for (i = 0; i < 4; i++)
if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
break;
*/
static void
convert_to_blend_type(struct gallivm_state *gallivm,
+ unsigned block_size,
const struct util_format_description *src_fmt,
struct lp_type src_type,
struct lp_type dst_type,
struct lp_type blend_type;
struct lp_type mem_type;
unsigned i, j, k;
- unsigned pixels = 16 / num_srcs;
+ unsigned pixels = block_size / num_srcs;
bool is_arith;
+ /*
+ * full custom path for packed floats and srgb formats - none of the later
+ * functions would do anything useful, and given the lp_type representation they
+ * can't be fixed. Should really have some SoA blend path for these kind of
+ * formats rather than hacking them in here.
+ */
+ if (format_expands_to_float_soa(src_fmt)) {
+ LLVMValueRef tmpsrc[4];
+ /*
+ * This is pretty suboptimal for this case blending in SoA would be much
+ * better, since conversion gets us SoA values so need to convert back.
+ */
+ assert(src_type.width == 32);
+ assert(dst_type.floating);
+ assert(dst_type.width == 32);
+ assert(dst_type.length % 4 == 0);
+ assert(num_srcs % 4 == 0);
+
+ for (i = 0; i < 4; i++) {
+ tmpsrc[i] = src[i];
+ }
+ for (i = 0; i < num_srcs / 4; i++) {
+ LLVMValueRef tmpsoa[4];
+ LLVMValueRef tmps = tmpsrc[i];
+ if (dst_type.length == 8) {
+ LLVMValueRef shuffles[8];
+ unsigned j;
+ /* fetch was 4 values but need 8-wide output values */
+ tmps = lp_build_concat(gallivm, &tmpsrc[i * 2], src_type, 2);
+ /*
+ * for 8-wide aos transpose would give us wrong order not matching
+ * incoming converted fs values and mask. ARGH.
+ */
+ for (j = 0; j < 4; j++) {
+ shuffles[j] = lp_build_const_int32(gallivm, j * 2);
+ shuffles[j + 4] = lp_build_const_int32(gallivm, j * 2 + 1);
+ }
+ tmps = LLVMBuildShuffleVector(builder, tmps, tmps,
+ LLVMConstVector(shuffles, 8), "");
+ }
+ if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
+ lp_build_r11g11b10_to_float(gallivm, tmps, tmpsoa);
+ }
+ else {
+ lp_build_unpack_rgba_soa(gallivm, src_fmt, dst_type, tmps, tmpsoa);
+ }
+ lp_build_transpose_aos(gallivm, dst_type, tmpsoa, &src[i * 4]);
+ }
+ return;
+ }
+
lp_mem_type_from_format_desc(src_fmt, &mem_type);
lp_blend_type_from_format_desc(src_fmt, &blend_type);
for (i = 0; i < num_srcs; ++i) {
LLVMValueRef chans[4];
- LLVMValueRef res;
- unsigned sa = 0;
+ LLVMValueRef res = NULL;
dst[i] = LLVMBuildZExt(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
for (j = 0; j < src_fmt->nr_channels; ++j) {
unsigned mask = 0;
+ unsigned sa = src_fmt->channel[j].shift;
+#ifdef PIPE_ARCH_LITTLE_ENDIAN
+ unsigned from_lsb = j;
+#else
+ unsigned from_lsb = src_fmt->nr_channels - j - 1;
+#endif
for (k = 0; k < src_fmt->channel[j].size; ++k) {
mask |= 1 << k;
/* Insert bits into correct position */
chans[j] = LLVMBuildShl(builder,
chans[j],
- lp_build_const_int_vec(gallivm, src_type, j * blend_type.width),
+ lp_build_const_int_vec(gallivm, src_type, from_lsb * blend_type.width),
"");
- sa += src_fmt->channel[j].size;
-
if (j == 0) {
res = chans[j];
} else {
*/
static void
convert_from_blend_type(struct gallivm_state *gallivm,
+ unsigned block_size,
const struct util_format_description *src_fmt,
struct lp_type src_type,
struct lp_type dst_type,
struct lp_type mem_type;
struct lp_type blend_type;
LLVMBuilderRef builder = gallivm->builder;
- unsigned pixels = 16 / num_srcs;
+ unsigned pixels = block_size / num_srcs;
bool is_arith;
+ /*
+ * full custom path for packed floats and srgb formats - none of the later
+ * functions would do anything useful, and given the lp_type representation they
+ * can't be fixed. Should really have some SoA blend path for these kind of
+ * formats rather than hacking them in here.
+ */
+ if (format_expands_to_float_soa(src_fmt)) {
+ /*
+ * This is pretty suboptimal for this case blending in SoA would be much
+ * better - we need to transpose the AoS values back to SoA values for
+ * conversion/packing.
+ */
+ assert(src_type.floating);
+ assert(src_type.width == 32);
+ assert(src_type.length % 4 == 0);
+ assert(dst_type.width == 32);
+
+ for (i = 0; i < num_srcs / 4; i++) {
+ LLVMValueRef tmpsoa[4], tmpdst;
+ lp_build_transpose_aos(gallivm, src_type, &src[i * 4], tmpsoa);
+ /* really really need SoA here */
+
+ if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
+ tmpdst = lp_build_float_to_r11g11b10(gallivm, tmpsoa);
+ }
+ else {
+ tmpdst = lp_build_float_to_srgb_packed(gallivm, src_fmt,
+ src_type, tmpsoa);
+ }
+
+ if (src_type.length == 8) {
+ LLVMValueRef tmpaos, shuffles[8];
+ unsigned j;
+ /*
+ * for 8-wide aos transpose has given us wrong order not matching
+ * output order. HMPF. Also need to split the output values manually.
+ */
+ for (j = 0; j < 4; j++) {
+ shuffles[j * 2] = lp_build_const_int32(gallivm, j);
+ shuffles[j * 2 + 1] = lp_build_const_int32(gallivm, j + 4);
+ }
+ tmpaos = LLVMBuildShuffleVector(builder, tmpdst, tmpdst,
+ LLVMConstVector(shuffles, 8), "");
+ src[i * 2] = lp_build_extract_range(gallivm, tmpaos, 0, 4);
+ src[i * 2 + 1] = lp_build_extract_range(gallivm, tmpaos, 4, 4);
+ }
+ else {
+ src[i] = tmpdst;
+ }
+ }
+ return;
+ }
+
lp_mem_type_from_format_desc(src_fmt, &mem_type);
lp_blend_type_from_format_desc(src_fmt, &blend_type);
for (i = 0; i < num_srcs; ++i) {
LLVMValueRef chans[4];
- LLVMValueRef res;
- unsigned sa = 0;
+ LLVMValueRef res = NULL;
dst[i] = LLVMBuildBitCast(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
for (j = 0; j < src_fmt->nr_channels; ++j) {
unsigned mask = 0;
+ unsigned sa = src_fmt->channel[j].shift;
+#ifdef PIPE_ARCH_LITTLE_ENDIAN
+ unsigned from_lsb = j;
+#else
+ unsigned from_lsb = src_fmt->nr_channels - j - 1;
+#endif
assert(blend_type.width > src_fmt->channel[j].size);
/* Extract bits */
chans[j] = LLVMBuildLShr(builder,
dst[i],
- lp_build_const_int_vec(gallivm, src_type, j * blend_type.width),
+ lp_build_const_int_vec(gallivm, src_type, from_lsb * blend_type.width),
"");
chans[j] = LLVMBuildAnd(builder,
row_type.length = alpha_type.length;
/* Twiddle the alpha to match pixels */
- lp_bld_quad_twiddle(gallivm, alpha_type, src_alpha, 4, src_alpha);
+ lp_bld_quad_twiddle(gallivm, alpha_type, src_alpha, block_height, src_alpha);
- for (i = 0; i < 4; ++i) {
+ /*
+ * TODO this should use single lp_build_conv call for
+ * src_count == 1 && dst_channels == 1 case (dropping the concat below)
+ */
+ for (i = 0; i < block_height; ++i) {
lp_build_conv(gallivm, alpha_type, row_type, &src_alpha[i], 1, &src_alpha[i], 1);
}
row_type.length = length;
/* If only one channel we can only need the single alpha value per pixel */
- if (src_count == 1) {
- assert(dst_channels == 1);
+ if (src_count == 1 && dst_channels == 1) {
- lp_build_concat_n(gallivm, alpha_type, src_alpha, 4, src_alpha, src_count);
+ lp_build_concat_n(gallivm, alpha_type, src_alpha, block_height, src_alpha, src_count);
} else {
/* If there are more srcs than rows then we need to split alpha up */
if (src_count > block_height) {
unsigned pixels = block_size / src_count;
unsigned idx = i - 1;
- src_alpha[idx] = lp_build_extract_range(gallivm, src_alpha[(idx * pixels) / 4], (idx * pixels) % 4, pixels);
+ src_alpha[idx] = lp_build_extract_range(gallivm, src_alpha[(idx * pixels) / 4],
+ (idx * pixels) % 4, pixels);
}
}
boolean do_branch)
{
const unsigned alpha_channel = 3;
- const unsigned block_width = 4;
- const unsigned block_height = 4;
+ const unsigned block_width = LP_RASTER_BLOCK_SIZE;
+ const unsigned block_height = LP_RASTER_BLOCK_SIZE;
const unsigned block_size = block_width * block_height;
const unsigned lp_integer_vector_width = 128;
LLVMValueRef blend_alpha;
LLVMValueRef i32_zero;
LLVMValueRef check_mask;
+ LLVMValueRef undef_src_val;
struct lp_build_mask_context mask_ctx;
struct lp_type mask_type;
const boolean dual_source_blend = variant->key.blend.rt[0].blend_enable &&
util_blend_state_is_dual(&variant->key.blend, 0);
+ const boolean is_1d = variant->key.resource_1d;
+ unsigned num_fullblock_fs = is_1d ? 2 * num_fs : num_fs;
+
mask_type = lp_int32_vec4_type();
mask_type.length = fs_type.length;
- /* Compute the alignment of the destination pointer in bytes */
-#if 0
- dst_alignment = (block_width * out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
-#else
- /* FIXME -- currently we're fetching pixels one by one, instead of row by row */
- dst_alignment = (1 * out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
-#endif
- /* Force power-of-two alignment by extracting only the least-significant-bit */
- dst_alignment = 1 << (ffs(dst_alignment) - 1);
- /* Resource base and stride pointers are aligned to 16 bytes, so that's the maximum alignment we can guarantee */
- dst_alignment = MIN2(dst_alignment, 16);
+ for (i = num_fs; i < num_fullblock_fs; i++) {
+ fs_mask[i] = lp_build_zero(gallivm, mask_type);
+ }
/* Do not bother executing code when mask is empty.. */
if (do_branch) {
check_mask = LLVMConstNull(lp_build_int_vec_type(gallivm, mask_type));
- for (i = 0; i < num_fs; ++i) {
+ for (i = 0; i < num_fullblock_fs; ++i) {
check_mask = LLVMBuildOr(builder, check_mask, fs_mask[i], "");
}
partial_mask |= !variant->opaque;
i32_zero = lp_build_const_int32(gallivm, 0);
+#if HAVE_LLVM < 0x0302
+ /*
+ * undef triggers a crash in LLVMBuildTrunc in convert_from_blend_type in some
+ * cases (seen with r10g10b10a2, 128bit wide vectors) (only used for 1d case).
+ */
+ undef_src_val = lp_build_zero(gallivm, fs_type);
+#else
+ undef_src_val = lp_build_undef(gallivm, fs_type);
+#endif
+
+
/* Get type from output format */
lp_blend_type_from_format_desc(out_format_desc, &row_type);
lp_mem_type_from_format_desc(out_format_desc, &dst_type);
}
}
+ if (format_expands_to_float_soa(out_format_desc)) {
+ /*
+ * the code above can't work for layout_other
+ * for srgb it would sort of work but we short-circuit swizzles, etc.
+ * as that is done as part of unpack / pack.
+ */
+ dst_channels = 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
+ has_alpha = true;
+ swizzle[0] = 0;
+ swizzle[1] = 1;
+ swizzle[2] = 2;
+ swizzle[3] = 3;
+ pad_inline = true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
+ }
+
/* If 3 channels then pad to include alpha for 4 element transpose */
if (dst_channels == 3 && !has_alpha) {
for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
/*
* Load shader output
*/
- for (i = 0; i < num_fs; ++i) {
+ for (i = 0; i < num_fullblock_fs; ++i) {
/* Always load alpha for use in blending */
- LLVMValueRef alpha = LLVMBuildLoad(builder, fs_out_color[rt][alpha_channel][i], "");
+ LLVMValueRef alpha;
+ if (i < num_fs) {
+ alpha = LLVMBuildLoad(builder, fs_out_color[rt][alpha_channel][i], "");
+ }
+ else {
+ alpha = undef_src_val;
+ }
/* Load each channel */
for (j = 0; j < dst_channels; ++j) {
assert(swizzle[j] < 4);
- fs_src[i][j] = LLVMBuildLoad(builder, fs_out_color[rt][swizzle[j]][i], "");
+ if (i < num_fs) {
+ fs_src[i][j] = LLVMBuildLoad(builder, fs_out_color[rt][swizzle[j]][i], "");
+ }
+ else {
+ fs_src[i][j] = undef_src_val;
+ }
}
/* If 3 channels then pad to include alpha for 4 element transpose */
}
if (dual_source_blend) {
/* same as above except different src/dst, skip masks and comments... */
- for (i = 0; i < num_fs; ++i) {
- LLVMValueRef alpha = LLVMBuildLoad(builder, fs_out_color[1][alpha_channel][i], "");
+ for (i = 0; i < num_fullblock_fs; ++i) {
+ LLVMValueRef alpha;
+ if (i < num_fs) {
+ alpha = LLVMBuildLoad(builder, fs_out_color[1][alpha_channel][i], "");
+ }
+ else {
+ alpha = undef_src_val;
+ }
for (j = 0; j < dst_channels; ++j) {
assert(swizzle[j] < 4);
- fs_src1[i][j] = LLVMBuildLoad(builder, fs_out_color[1][swizzle[j]][i], "");
+ if (i < num_fs) {
+ fs_src1[i][j] = LLVMBuildLoad(builder, fs_out_color[1][swizzle[j]][i], "");
+ }
+ else {
+ fs_src1[i][j] = undef_src_val;
+ }
}
if (dst_channels == 3 && !has_alpha) {
fs_src1[i][3] = alpha;
*/
fs_type.floating = 0;
fs_type.sign = dst_type.sign;
- for (i = 0; i < num_fs; ++i) {
+ for (i = 0; i < num_fullblock_fs; ++i) {
for (j = 0; j < dst_channels; ++j) {
fs_src[i][j] = LLVMBuildBitCast(builder, fs_src[i][j],
lp_build_vec_type(gallivm, fs_type), "");
/*
* Pixel twiddle from fragment shader order to memory order
*/
- src_count = generate_fs_twiddle(gallivm, fs_type, num_fs, dst_channels, fs_src, src, pad_inline);
+ src_count = generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs,
+ dst_channels, fs_src, src, pad_inline);
if (dual_source_blend) {
- generate_fs_twiddle(gallivm, fs_type, num_fs, dst_channels, fs_src1, src1, pad_inline);
+ generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs, dst_channels,
+ fs_src1, src1, pad_inline);
}
src_channels = dst_channels < 3 ? dst_channels : 4;
- if (src_count != num_fs * src_channels) {
- unsigned ds = src_count / (num_fs * src_channels);
+ if (src_count != num_fullblock_fs * src_channels) {
+ unsigned ds = src_count / (num_fullblock_fs * src_channels);
row_type.length /= ds;
fs_type.length = row_type.length;
}
unsigned bits = row_type.width * row_type.length;
unsigned combined;
+ assert(src_count >= (vector_width / bits));
+
dst_count = src_count / (vector_width / bits);
+
combined = lp_build_concat_n(gallivm, row_type, src, src_count, src, dst_count);
if (dual_source_blend) {
lp_build_concat_n(gallivm, row_type, src1, src_count, src1, dst_count);
/*
* Mask conversion
*/
- lp_bld_quad_twiddle(gallivm, mask_type, &src_mask[0], 4, &src_mask[0]);
+ lp_bld_quad_twiddle(gallivm, mask_type, &src_mask[0], block_height, &src_mask[0]);
if (src_count < block_height) {
lp_build_concat_n(gallivm, mask_type, src_mask, 4, src_mask, src_count);
unsigned pixels = block_size / src_count;
unsigned idx = i - 1;
- src_mask[idx] = lp_build_extract_range(gallivm, src_mask[(idx * pixels) / 4], (idx * pixels) % 4, pixels);
+ src_mask[idx] = lp_build_extract_range(gallivm, src_mask[(idx * pixels) / 4],
+ (idx * pixels) % 4, pixels);
}
}
dst_count = src_count;
}
- dst_type.length *= 16 / dst_count;
+ dst_type.length *= block_size / dst_count;
+
+ if (format_expands_to_float_soa(out_format_desc)) {
+ /*
+ * we need multiple values at once for the conversion, so can as well
+ * load them vectorized here too instead of concatenating later.
+ * (Still need concatenation later for 8-wide vectors).
+ */
+ dst_count = block_height;
+ dst_type.length = block_width;
+ }
+
+ /*
+ * Compute the alignment of the destination pointer in bytes
+ * We fetch 1-4 pixels, if the format has pot alignment then those fetches
+ * are always aligned by MIN2(16, fetch_width) except for buffers (not
+ * 1d tex but can't distinguish here) so need to stick with per-pixel
+ * alignment in this case.
+ */
+ if (is_1d) {
+ dst_alignment = (out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
+ }
+ else {
+ dst_alignment = dst_type.length * dst_type.width / 8;
+ }
+ /* Force power-of-two alignment by extracting only the least-significant-bit */
+ dst_alignment = 1 << (ffs(dst_alignment) - 1);
+ /*
+ * Resource base and stride pointers are aligned to 16 bytes, so that's
+ * the maximum alignment we can guarantee
+ */
+ dst_alignment = MIN2(16, dst_alignment);
+
+ if (is_1d) {
+ load_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,
+ dst, dst_type, dst_count / 4, dst_alignment);
+ for (i = dst_count / 4; i < dst_count; i++) {
+ dst[i] = lp_build_undef(gallivm, dst_type);
+ }
- load_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
- dst, dst_type, dst_count, dst_alignment);
+ }
+ else {
+ load_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
+ dst, dst_type, dst_count, dst_alignment);
+ }
/*
* It seems some cleanup could be done here (like skipping conversion/blend
* when not needed).
*/
- convert_to_blend_type(gallivm, out_format_desc, dst_type, row_type, dst, src_count);
+ convert_to_blend_type(gallivm, block_size, out_format_desc, dst_type, row_type, dst, src_count);
for (i = 0; i < src_count; ++i) {
dst[i] = lp_build_blend_aos(gallivm,
pad_inline ? 4 : dst_channels);
}
- convert_from_blend_type(gallivm, out_format_desc, row_type, dst_type, dst, src_count);
+ convert_from_blend_type(gallivm, block_size, out_format_desc, row_type, dst_type, dst, src_count);
/* Split the blend rows back to memory rows */
if (dst_count > src_count) {
src_count *= 2;
}
-
/*
* Store blend result to memory
*/
- store_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
- dst, dst_type, dst_count, dst_alignment);
+ if (is_1d) {
+ store_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,
+ dst, dst_type, dst_count / 4, dst_alignment);
+ }
+ else {
+ store_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
+ dst, dst_type, dst_count, dst_alignment);
+ }
if (do_branch) {
lp_build_mask_end(&mask_ctx);
struct lp_type blend_type;
LLVMTypeRef fs_elem_type;
LLVMTypeRef blend_vec_type;
- LLVMTypeRef arg_types[12];
+ LLVMTypeRef arg_types[13];
LLVMTypeRef func_type;
LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
LLVMValueRef color_ptr_ptr;
LLVMValueRef stride_ptr;
LLVMValueRef depth_ptr;
+ LLVMValueRef depth_stride;
LLVMValueRef mask_input;
LLVMValueRef thread_data_ptr;
LLVMBasicBlockRef block;
LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][16 / 4];
LLVMValueRef function;
LLVMValueRef facing;
- const struct util_format_description *zs_format_desc;
unsigned num_fs;
unsigned i;
unsigned chan;
unsigned cbuf;
boolean cbuf0_write_all;
- boolean try_loop = TRUE;
const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
util_blend_state_is_dual(&key->blend, 0);
fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
fs_type.width = 32; /* 32-bit float */
fs_type.length = MIN2(lp_native_vector_width / 32, 16); /* n*4 elements per vector */
- num_fs = 16 / fs_type.length; /* number of loops per 4x4 stamp */
memset(&blend_type, 0, sizeof blend_type);
blend_type.floating = FALSE; /* values are integers */
arg_types[9] = int32_type; /* mask_input */
arg_types[10] = variant->jit_thread_data_ptr_type; /* per thread data */
arg_types[11] = LLVMPointerType(int32_type, 0); /* stride */
+ arg_types[12] = int32_type; /* depth_stride */
func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
arg_types, Elements(arg_types), 0);
mask_input = LLVMGetParam(function, 9);
thread_data_ptr = LLVMGetParam(function, 10);
stride_ptr = LLVMGetParam(function, 11);
+ depth_stride = LLVMGetParam(function, 12);
lp_build_name(context_ptr, "context");
lp_build_name(x, "x");
lp_build_name(thread_data_ptr, "thread_data");
lp_build_name(mask_input, "mask_input");
lp_build_name(stride_ptr, "stride_ptr");
+ lp_build_name(depth_stride, "depth_stride");
/*
* Function body
/* code generated texture sampling */
sampler = lp_llvm_sampler_soa_create(key->state, context_ptr);
- zs_format_desc = util_format_description(key->zsbuf_format);
-
- if (!try_loop) {
- /*
- * The shader input interpolation info is not explicitely baked in the
- * shader key, but everything it derives from (TGSI, and flatshade) is
- * already included in the shader key.
- */
- lp_build_interp_soa_init(&interp,
- gallivm,
- shader->info.base.num_inputs,
- inputs,
- builder, fs_type,
- FALSE,
- a0_ptr, dadx_ptr, dady_ptr,
- x, y);
-
- /* loop over quads in the block */
- for(i = 0; i < num_fs; ++i) {
- LLVMValueRef depth_offset = LLVMConstInt(int32_type,
- i*fs_type.length*zs_format_desc->block.bits/8,
- 0);
- LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
- LLVMValueRef depth_ptr_i;
-
- depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &depth_offset, 1, "");
-
- generate_fs(gallivm,
- shader, key,
- builder,
- fs_type,
- context_ptr,
- i,
- &interp,
- sampler,
- &fs_mask[i], /* output */
- out_color,
- depth_ptr_i,
- facing,
- partial_mask,
- mask_input,
- thread_data_ptr);
+ num_fs = 16 / fs_type.length; /* number of loops per 4x4 stamp */
+ /* for 1d resources only run "upper half" of stamp */
+ if (key->resource_1d)
+ num_fs /= 2;
- for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
- for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan)
- fs_out_color[cbuf][chan][i] =
- out_color[cbuf * !cbuf0_write_all][chan];
- }
- if (dual_source_blend) {
- /* only support one dual source blend target hence always use output 1 */
- for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan)
- fs_out_color[1][chan][i] =
- out_color[1][chan];
- }
- }
- }
- else {
- unsigned depth_bits = zs_format_desc->block.bits/8;
+ {
LLVMValueRef num_loop = lp_build_const_int32(gallivm, num_fs);
LLVMTypeRef mask_type = lp_build_int_vec_type(gallivm, fs_type);
LLVMValueRef mask_store = lp_build_array_alloca(gallivm, mask_type,
gallivm,
shader->info.base.num_inputs,
inputs,
+ shader->info.base.pixel_center_integer,
builder, fs_type,
- TRUE,
a0_ptr, dadx_ptr, dady_ptr,
x, y);
mask_store, /* output */
color_store,
depth_ptr,
- depth_bits,
+ depth_stride,
facing,
thread_data_ptr);
memset(key, 0, shader->variant_key_size);
if (lp->framebuffer.zsbuf) {
- if (lp->depth_stencil->depth.enabled) {
- key->zsbuf_format = lp->framebuffer.zsbuf->format;
+ enum pipe_format zsbuf_format = lp->framebuffer.zsbuf->format;
+ const struct util_format_description *zsbuf_desc =
+ util_format_description(zsbuf_format);
+
+ if (lp->depth_stencil->depth.enabled &&
+ util_format_has_depth(zsbuf_desc)) {
+ key->zsbuf_format = zsbuf_format;
memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
}
- if (lp->depth_stencil->stencil[0].enabled) {
- key->zsbuf_format = lp->framebuffer.zsbuf->format;
+ if (lp->depth_stencil->stencil[0].enabled &&
+ util_format_has_stencil(zsbuf_desc)) {
+ key->zsbuf_format = zsbuf_format;
memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
}
+ if (llvmpipe_resource_is_1d(lp->framebuffer.zsbuf->texture)) {
+ key->resource_1d = TRUE;
+ }
}
/* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
/* alpha.ref_value is passed in jit_context */
key->flatshade = lp->rasterizer->flatshade;
- if (lp->active_occlusion_query) {
+ if (lp->active_occlusion_queries) {
key->occlusion_count = TRUE;
}
key->cbuf_format[i] = format;
+ /*
+ * Figure out if this is a 1d resource. Note that OpenGL allows crazy
+ * mixing of 2d textures with height 1 and 1d textures, so make sure
+ * we pick 1d if any cbuf or zsbuf is 1d.
+ */
+ if (llvmpipe_resource_is_1d(lp->framebuffer.cbufs[0]->texture)) {
+ key->resource_1d = TRUE;
+ }
+
format_desc = util_format_description(format);
assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);