#include "midgard/midgard_compile.h"
#include "compiler/nir/nir_builder.h"
#include "nir/nir_lower_blend.h"
+#include "panfrost/util/pan_lower_framebuffer.h"
#include "gallium/auxiliary/util/u_blend.h"
#include "util/u_memory.h"
*
* As is typical with Midgard, shader binaries must be accompanied by
* information about the first tag (ORed with the bottom nibble of address,
- * like usual) and work registers. Work register count is specified in the
- * blend descriptor, as well as in the coresponding fragment shader's work
- * count. This suggests that blend shader invocation is tied to fragment shader
+ * like usual) and work registers. Work register count is assumed to be less
+ * than or equal to the coresponding fragment shader's work count. This
+ * suggests that blend shader invocation is tied to fragment shader
* execution.
*
- * ---
- *
- * As for blend shaders, they use the standard ISA.
- *
- * The source pixel colour, including alpha, is preloaded into r0 as a vec4 of
- * float32.
- *
- * The destination pixel colour must be loaded explicitly via load/store ops.
- * TODO: Investigate.
- *
- * They use fragment shader writeout; however, instead of writing a vec4 of
- * float32 for RGBA encoding, we writeout a vec4 of uint8, using 8-bit imov
- * instead of 32-bit fmov. The net result is that r0 encodes a single uint32
- * containing all four channels of the color. Accordingly, the blend shader
- * epilogue has to scale all four channels by 255 and then type convert to a
- * uint8.
- *
- * ---
+ * The shaders themselves use the standard ISA. The source pixel colour,
+ * including alpha, is preloaded into r0 as a vec4 of float32. The destination
+ * pixel colour must be loaded explicitly via load/store ops, possibly
+ * performing conversions in software. The blended colour must be stored with a
+ * fragment writeout in the correct framebuffer format, either in software or
+ * via conversion opcodes on the load/store pipe.
*
* Blend shaders hardcode constants. Naively, this requires recompilation each
* time the blend color changes, which is a performance risk. Accordingly, we
static nir_lower_blend_options
nir_make_options(const struct pipe_blend_state *blend, unsigned i)
{
- nir_lower_blend_options options;
+ nir_lower_blend_options options = { 0 };
if (blend->logicop_enable) {
options.logicop_enable = true;
options.logicop_enable = false;
+ if (!blend->independent_blend_enable)
+ i = 0;
+
/* If blend is disabled, we just use replace mode */
nir_lower_blend_channel rgb = {
return options;
}
+static nir_ssa_def *
+nir_iclamp(nir_builder *b, nir_ssa_def *v, int32_t lo, int32_t hi)
+{
+ return nir_imin(b, nir_imax(b, v, nir_imm_int(b, lo)), nir_imm_int(b, hi));
+}
+
struct panfrost_blend_shader
panfrost_compile_blend_shader(
struct panfrost_context *ctx,
nir_function *fn = nir_function_create(shader, "main");
nir_function_impl *impl = nir_function_impl_create(fn);
+ const struct util_format_description *format_desc =
+ util_format_description(format);
+
+ nir_alu_type T = pan_unpacked_type_for_format(format_desc);
+ enum glsl_base_type g =
+ (T == nir_type_float16) ? GLSL_TYPE_FLOAT16 :
+ (T == nir_type_float32) ? GLSL_TYPE_FLOAT :
+ (T == nir_type_int8) ? GLSL_TYPE_INT8 :
+ (T == nir_type_int16) ? GLSL_TYPE_INT16 :
+ (T == nir_type_int32) ? GLSL_TYPE_INT :
+ (T == nir_type_uint8) ? GLSL_TYPE_UINT8 :
+ (T == nir_type_uint16) ? GLSL_TYPE_UINT16 :
+ (T == nir_type_uint32) ? GLSL_TYPE_UINT :
+ GLSL_TYPE_FLOAT;
+
/* Create the blend variables */
nir_variable *c_src = nir_variable_create(shader, nir_var_shader_in, glsl_vector_type(GLSL_TYPE_FLOAT, 4), "gl_Color");
- nir_variable *c_out = nir_variable_create(shader, nir_var_shader_out, glsl_vector_type(GLSL_TYPE_FLOAT, 4), "gl_FragColor");
+ nir_variable *c_src1 = nir_variable_create(shader, nir_var_shader_in, glsl_vector_type(GLSL_TYPE_FLOAT, 4), "gl_Color1");
+ nir_variable *c_out = nir_variable_create(shader, nir_var_shader_out, glsl_vector_type(g, 4), "gl_FragColor");
c_src->data.location = VARYING_SLOT_COL0;
+ c_src1->data.location = VARYING_SLOT_VAR0;
c_out->data.location = FRAG_RESULT_COLOR;
+ c_src1->data.driver_location = 1;
+
/* Setup nir_builder */
nir_builder _b;
/* Setup inputs */
- nir_ssa_def *s_src = nir_load_var(b, c_src);
+ nir_ssa_def *s_src[] = {nir_load_var(b, c_src), nir_load_var(b, c_src1)};
+
+ for (int i = 0; i < ARRAY_SIZE(s_src); ++i) {
+ if (T == nir_type_float16)
+ s_src[i] = nir_f2f16(b, s_src[i]);
+ else if (T == nir_type_int16)
+ s_src[i] = nir_i2i16(b, nir_iclamp(b, s_src[i], -32768, 32767));
+ else if (T == nir_type_uint16)
+ s_src[i] = nir_u2u16(b, nir_umin(b, s_src[i], nir_imm_int(b, 65535)));
+ else if (T == nir_type_int8)
+ s_src[i] = nir_i2i8(b, nir_iclamp(b, s_src[i], -128, 127));
+ else if (T == nir_type_uint8)
+ s_src[i] = nir_u2u8(b, nir_umin(b, s_src[i], nir_imm_int(b, 255)));
+ }
/* Build a trivial blend shader */
- nir_store_var(b, c_out, s_src, 0xFF);
+ nir_store_var(b, c_out, s_src[0], 0xFF);
nir_lower_blend_options options =
nir_make_options(cso, rt);
options.format = format;
+ options.src1 = s_src[1];
- NIR_PASS_V(shader, nir_lower_blend, options);
+ if (T == nir_type_float16)
+ options.half = true;
- NIR_PASS_V(shader, nir_lower_framebuffer, format, dev->gpu_id);
+ NIR_PASS_V(shader, nir_lower_blend, options);
/* Compile the built shader */
- panfrost_program program;
- midgard_compile_shader_nir(shader, &program, true, rt, dev->gpu_id, false);
+ panfrost_program program = {
+ .rt_formats = {format}
+ };
+
+ midgard_compile_shader_nir(shader, &program, true, rt, dev->gpu_id, false, false);
/* Allow us to patch later */
res.patch_index = program.blend_patch_offset;
res.first_tag = program.first_tag;
res.size = program.compiled.size;
res.buffer = program.compiled.data;
+ res.work_count = program.work_register_count;
return res;
}