wu-arch/vortex
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fixes: texture unit mem access sometimes going to smem, bilinear texture filtering; new: cache req_id,

Browse Source
This commit is contained in:
Blaise Tine
2021-11-24 00:00:17 -05:00
parent 1501360f4b
commit 18762dffce
70 changed files with 3818 additions and 1727 deletions
Download Patch File Download Diff File Expand all files Collapse all files

73
tests/regression/tex/kernel.c
View File

@@ -1,11 +1,9 @@
#include <stdint.h>
#include <vx_intrinsics.h>
#include <vx_spawn.h>
#include "common.h"
#include <vx_print.h>
#include "texsw.h"
#define ENABLE_SW
typedef struct {
kernel_arg_t* state;
uint32_t tile_width;
@@ -14,29 +12,50 @@ typedef struct {
float deltaY;
} tile_arg_t;
template <typename T, T Start, T End>
struct static_for_t {
template <typename Fn>
inline void operator()(const Fn& callback) const {
callback(Start);
static_for_t<T, Start+1, End>()(callback);
}
};
template <typename T, T N>
struct static_for_t<T, N, N> {
template <typename Fn>
inline void operator()(const Fn& callback) const {}
};
void kernel_body(int task_id, tile_arg_t* arg) {
kernel_arg_t* state = arg->state;
uint32_t xoffset = 0;
uint32_t yoffset = task_id * arg->tile_height;
uint8_t* dst_ptr = (uint8_t*)(state->dst_ptr + xoffset * state->dst_stride + yoffset * state->dst_pitch);
uint32_t yoffset = task_id * arg->tile_height;
float fv = yoffset * arg->deltaY;
uint8_t* dst_ptr = (uint8_t*)(state->dst_addr + xoffset * state->dst_stride + yoffset * state->dst_pitch);
Fixed<16> xlod(state->lod);
/*vx_printf("task_id=%d, deltaX=%f, deltaY=%f, tile_width=%d, tile_height=%d\n",
task_id, arg->deltaX, arg->deltaY, arg->tile_width, arg->tile_height);*/
float fv = (yoffset + 0.5f) * arg->deltaY;
for (uint32_t y = 0; y < arg->tile_height; ++y) {
uint32_t* dst_row = (uint32_t*)dst_ptr;
float fu = xoffset * arg->deltaX;
float fu = (xoffset + 0.5f) * arg->deltaX;
for (uint32_t x = 0; x < arg->tile_width; ++x) {
int32_t u = (int32_t)(fu * (1<<20));
int32_t v = (int32_t)(fv * (1<<20));
Fixed<TEX_FXD_FRAC> xu(fu);
Fixed<TEX_FXD_FRAC> xv(fv);
uint32_t color;
#ifdef ENABLE_SW
if (state->use_sw) {
dst_row[x] = (state->filter == 2) ? tex3_sw(state, 0, u, v, state->lod) : tex_sw(state, 0, u, v, state->lod);
} else {
#endif
dst_row[x] = (state->filter == 2) ? vx_tex3(0, u, v, state->lod) : vx_tex(0, u, v, state->lod);
#ifdef ENABLE_SW
}
if (state->use_sw)
color = tex_load_sw(state, xu, xv, xlod);
else
#endif
color = tex_load_hw(state, xu, xv, xlod);
//vx_printf("task_id=%d, x=%d, y=%d, fu=%f, fv=%f, xu=0x%x, xv=0x%x, color=0x%x\n", task_id, x, y, fu, fv, xu.data(), xv.data(), color);
dst_row[x] = color;
fu += arg->deltaX;
}
dst_ptr += state->dst_pitch;
@@ -48,13 +67,16 @@ int main() {
kernel_arg_t* arg = (kernel_arg_t*)KERNEL_ARG_DEV_MEM_ADDR;
// configure texture unit
vx_csr_write(CSR_TEX_ADDR(0), arg->src_ptr);
vx_csr_write(CSR_TEX_MIPOFF(0), 0);
vx_csr_write(CSR_TEX_WIDTH(0), arg->src_logWidth);
vx_csr_write(CSR_TEX_HEIGHT(0), arg->src_logHeight);
vx_csr_write(CSR_TEX_FORMAT(0), arg->format);
vx_csr_write(CSR_TEX_WRAP(0), (arg->wrap << 2) | arg->wrap);
vx_csr_write(CSR_TEX_FILTER(0), (arg->filter ? 1 : 0));
csr_write(CSR_TEX(0, TEX_STATE_WIDTH), arg->src_logwidth);
csr_write(CSR_TEX(0, TEX_STATE_HEIGHT), arg->src_logheight);
csr_write(CSR_TEX(0, TEX_STATE_FORMAT), arg->format);
csr_write(CSR_TEX(0, TEX_STATE_WRAPU), arg->wrapu);
csr_write(CSR_TEX(0, TEX_STATE_WRAPV), arg->wrapv);
csr_write(CSR_TEX(0, TEX_STATE_FILTER), (arg->filter ? 1 : 0));
csr_write(CSR_TEX(0, TEX_STATE_ADDR), arg->src_addr);
static_for_t<int, 0, TEX_LOD_MAX+1>()([&](int i) {
csr_write(CSR_TEX(0, TEX_STATE_MIPOFF(i)), arg->mip_offs[i]);
});
tile_arg_t targ;
targ.state = arg;
@@ -64,4 +86,9 @@ int main() {
targ.deltaY = 1.0f / arg->dst_height;
vx_spawn_tasks(arg->num_tasks, (vx_spawn_tasks_cb)kernel_body, &targ);
/*for (uint32_t t=0; t < arg->num_tasks; ++t) {
kernel_body(t, &targ);
}*/
return 0;
}